2 * Copyright © 2017 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included
12 * in all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20 * DEALINGS IN THE SOFTWARE.
26 * ============================= GENXML CODE =============================
27 * [This file is compiled once per generation.]
28 * =======================================================================
30 * This is the main state upload code.
32 * Gallium uses Constant State Objects, or CSOs, for most state. Large,
33 * complex, or highly reusable state can be created once, and bound and
34 * rebound multiple times. This is modeled with the pipe->create_*_state()
35 * and pipe->bind_*_state() hooks. Highly dynamic or inexpensive state is
36 * streamed out on the fly, via pipe->set_*_state() hooks.
38 * OpenGL involves frequently mutating context state, which is mirrored in
39 * core Mesa by highly mutable data structures. However, most applications
40 * typically draw the same things over and over - from frame to frame, most
41 * of the same objects are still visible and need to be redrawn. So, rather
42 * than inventing new state all the time, applications usually mutate to swap
43 * between known states that we've seen before.
45 * Gallium isolates us from this mutation by tracking API state, and
46 * distilling it into a set of Constant State Objects, or CSOs. Large,
47 * complex, or typically reusable state can be created once, then reused
48 * multiple times. Drivers can create and store their own associated data.
49 * This create/bind model corresponds to the pipe->create_*_state() and
50 * pipe->bind_*_state() driver hooks.
52 * Some state is cheap to create, or expected to be highly dynamic. Rather
53 * than creating and caching piles of CSOs for these, Gallium simply streams
54 * them out, via the pipe->set_*_state() driver hooks.
56 * To reduce draw time overhead, we try to compute as much state at create
57 * time as possible. Wherever possible, we translate the Gallium pipe state
58 * to 3DSTATE commands, and store those commands in the CSO. At draw time,
59 * we can simply memcpy them into a batch buffer.
61 * No hardware matches the abstraction perfectly, so some commands require
62 * information from multiple CSOs. In this case, we can store two copies
63 * of the packet (one in each CSO), and simply | together their DWords at
64 * draw time. Sometimes the second set is trivial (one or two fields), so
65 * we simply pack it at draw time.
67 * There are two main components in the file below. First, the CSO hooks
68 * create/bind/track state. The second are the draw-time upload functions,
69 * iris_upload_render_state() and iris_upload_compute_state(), which read
70 * the context state and emit the commands into the actual batch.
81 #define __gen_validate_value(x) VALGRIND_CHECK_MEM_IS_DEFINED(&(x), sizeof(x))
87 #include "pipe/p_defines.h"
88 #include "pipe/p_state.h"
89 #include "pipe/p_context.h"
90 #include "pipe/p_screen.h"
91 #include "util/u_inlines.h"
92 #include "util/u_format.h"
93 #include "util/u_framebuffer.h"
94 #include "util/u_transfer.h"
95 #include "util/u_upload_mgr.h"
96 #include "util/u_viewport.h"
99 #include "intel/compiler/brw_compiler.h"
100 #include "intel/common/gen_l3_config.h"
101 #include "intel/common/gen_sample_positions.h"
102 #include "iris_batch.h"
103 #include "iris_context.h"
104 #include "iris_pipe.h"
105 #include "iris_resource.h"
107 #define __gen_address_type struct iris_address
108 #define __gen_user_data struct iris_batch
110 #define ARRAY_BYTES(x) (sizeof(uint32_t) * ARRAY_SIZE(x))
113 __gen_combine_address(struct iris_batch
*batch
, void *location
,
114 struct iris_address addr
, uint32_t delta
)
116 uint64_t result
= addr
.offset
+ delta
;
119 iris_use_pinned_bo(batch
, addr
.bo
, addr
.write
);
120 /* Assume this is a general address, not relative to a base. */
121 result
+= addr
.bo
->gtt_offset
;
127 #define __genxml_cmd_length(cmd) cmd ## _length
128 #define __genxml_cmd_length_bias(cmd) cmd ## _length_bias
129 #define __genxml_cmd_header(cmd) cmd ## _header
130 #define __genxml_cmd_pack(cmd) cmd ## _pack
132 #define _iris_pack_command(batch, cmd, dst, name) \
133 for (struct cmd name = { __genxml_cmd_header(cmd) }, \
134 *_dst = (void *)(dst); __builtin_expect(_dst != NULL, 1); \
135 ({ __genxml_cmd_pack(cmd)(batch, (void *)_dst, &name); \
139 #define iris_pack_command(cmd, dst, name) \
140 _iris_pack_command(NULL, cmd, dst, name)
142 #define iris_pack_state(cmd, dst, name) \
143 for (struct cmd name = {}, \
144 *_dst = (void *)(dst); __builtin_expect(_dst != NULL, 1); \
145 __genxml_cmd_pack(cmd)(NULL, (void *)_dst, &name), \
148 #define iris_emit_cmd(batch, cmd, name) \
149 _iris_pack_command(batch, cmd, iris_get_command_space(batch, 4 * __genxml_cmd_length(cmd)), name)
151 #define iris_emit_merge(batch, dwords0, dwords1, num_dwords) \
153 uint32_t *dw = iris_get_command_space(batch, 4 * num_dwords); \
154 for (uint32_t i = 0; i < num_dwords; i++) \
155 dw[i] = (dwords0)[i] | (dwords1)[i]; \
156 VG(VALGRIND_CHECK_MEM_IS_DEFINED(dw, num_dwords)); \
159 #include "genxml/genX_pack.h"
160 #include "genxml/gen_macros.h"
161 #include "genxml/genX_bits.h"
163 #define MOCS_WB (2 << 1)
166 * Statically assert that PIPE_* enums match the hardware packets.
167 * (As long as they match, we don't need to translate them.)
169 UNUSED
static void pipe_asserts()
171 #define PIPE_ASSERT(x) STATIC_ASSERT((int)x)
173 /* pipe_logicop happens to match the hardware. */
174 PIPE_ASSERT(PIPE_LOGICOP_CLEAR
== LOGICOP_CLEAR
);
175 PIPE_ASSERT(PIPE_LOGICOP_NOR
== LOGICOP_NOR
);
176 PIPE_ASSERT(PIPE_LOGICOP_AND_INVERTED
== LOGICOP_AND_INVERTED
);
177 PIPE_ASSERT(PIPE_LOGICOP_COPY_INVERTED
== LOGICOP_COPY_INVERTED
);
178 PIPE_ASSERT(PIPE_LOGICOP_AND_REVERSE
== LOGICOP_AND_REVERSE
);
179 PIPE_ASSERT(PIPE_LOGICOP_INVERT
== LOGICOP_INVERT
);
180 PIPE_ASSERT(PIPE_LOGICOP_XOR
== LOGICOP_XOR
);
181 PIPE_ASSERT(PIPE_LOGICOP_NAND
== LOGICOP_NAND
);
182 PIPE_ASSERT(PIPE_LOGICOP_AND
== LOGICOP_AND
);
183 PIPE_ASSERT(PIPE_LOGICOP_EQUIV
== LOGICOP_EQUIV
);
184 PIPE_ASSERT(PIPE_LOGICOP_NOOP
== LOGICOP_NOOP
);
185 PIPE_ASSERT(PIPE_LOGICOP_OR_INVERTED
== LOGICOP_OR_INVERTED
);
186 PIPE_ASSERT(PIPE_LOGICOP_COPY
== LOGICOP_COPY
);
187 PIPE_ASSERT(PIPE_LOGICOP_OR_REVERSE
== LOGICOP_OR_REVERSE
);
188 PIPE_ASSERT(PIPE_LOGICOP_OR
== LOGICOP_OR
);
189 PIPE_ASSERT(PIPE_LOGICOP_SET
== LOGICOP_SET
);
191 /* pipe_blend_func happens to match the hardware. */
192 PIPE_ASSERT(PIPE_BLENDFACTOR_ONE
== BLENDFACTOR_ONE
);
193 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_COLOR
== BLENDFACTOR_SRC_COLOR
);
194 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA
== BLENDFACTOR_SRC_ALPHA
);
195 PIPE_ASSERT(PIPE_BLENDFACTOR_DST_ALPHA
== BLENDFACTOR_DST_ALPHA
);
196 PIPE_ASSERT(PIPE_BLENDFACTOR_DST_COLOR
== BLENDFACTOR_DST_COLOR
);
197 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
== BLENDFACTOR_SRC_ALPHA_SATURATE
);
198 PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_COLOR
== BLENDFACTOR_CONST_COLOR
);
199 PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_ALPHA
== BLENDFACTOR_CONST_ALPHA
);
200 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_COLOR
== BLENDFACTOR_SRC1_COLOR
);
201 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_ALPHA
== BLENDFACTOR_SRC1_ALPHA
);
202 PIPE_ASSERT(PIPE_BLENDFACTOR_ZERO
== BLENDFACTOR_ZERO
);
203 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_COLOR
== BLENDFACTOR_INV_SRC_COLOR
);
204 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_ALPHA
== BLENDFACTOR_INV_SRC_ALPHA
);
205 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_ALPHA
== BLENDFACTOR_INV_DST_ALPHA
);
206 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_COLOR
== BLENDFACTOR_INV_DST_COLOR
);
207 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_COLOR
== BLENDFACTOR_INV_CONST_COLOR
);
208 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_ALPHA
== BLENDFACTOR_INV_CONST_ALPHA
);
209 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_COLOR
== BLENDFACTOR_INV_SRC1_COLOR
);
210 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_ALPHA
== BLENDFACTOR_INV_SRC1_ALPHA
);
212 /* pipe_blend_func happens to match the hardware. */
213 PIPE_ASSERT(PIPE_BLEND_ADD
== BLENDFUNCTION_ADD
);
214 PIPE_ASSERT(PIPE_BLEND_SUBTRACT
== BLENDFUNCTION_SUBTRACT
);
215 PIPE_ASSERT(PIPE_BLEND_REVERSE_SUBTRACT
== BLENDFUNCTION_REVERSE_SUBTRACT
);
216 PIPE_ASSERT(PIPE_BLEND_MIN
== BLENDFUNCTION_MIN
);
217 PIPE_ASSERT(PIPE_BLEND_MAX
== BLENDFUNCTION_MAX
);
219 /* pipe_stencil_op happens to match the hardware. */
220 PIPE_ASSERT(PIPE_STENCIL_OP_KEEP
== STENCILOP_KEEP
);
221 PIPE_ASSERT(PIPE_STENCIL_OP_ZERO
== STENCILOP_ZERO
);
222 PIPE_ASSERT(PIPE_STENCIL_OP_REPLACE
== STENCILOP_REPLACE
);
223 PIPE_ASSERT(PIPE_STENCIL_OP_INCR
== STENCILOP_INCRSAT
);
224 PIPE_ASSERT(PIPE_STENCIL_OP_DECR
== STENCILOP_DECRSAT
);
225 PIPE_ASSERT(PIPE_STENCIL_OP_INCR_WRAP
== STENCILOP_INCR
);
226 PIPE_ASSERT(PIPE_STENCIL_OP_DECR_WRAP
== STENCILOP_DECR
);
227 PIPE_ASSERT(PIPE_STENCIL_OP_INVERT
== STENCILOP_INVERT
);
229 /* pipe_sprite_coord_mode happens to match 3DSTATE_SBE */
230 PIPE_ASSERT(PIPE_SPRITE_COORD_UPPER_LEFT
== UPPERLEFT
);
231 PIPE_ASSERT(PIPE_SPRITE_COORD_LOWER_LEFT
== LOWERLEFT
);
236 translate_prim_type(enum pipe_prim_type prim
, uint8_t verts_per_patch
)
238 static const unsigned map
[] = {
239 [PIPE_PRIM_POINTS
] = _3DPRIM_POINTLIST
,
240 [PIPE_PRIM_LINES
] = _3DPRIM_LINELIST
,
241 [PIPE_PRIM_LINE_LOOP
] = _3DPRIM_LINELOOP
,
242 [PIPE_PRIM_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
243 [PIPE_PRIM_TRIANGLES
] = _3DPRIM_TRILIST
,
244 [PIPE_PRIM_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
245 [PIPE_PRIM_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
246 [PIPE_PRIM_QUADS
] = _3DPRIM_QUADLIST
,
247 [PIPE_PRIM_QUAD_STRIP
] = _3DPRIM_QUADSTRIP
,
248 [PIPE_PRIM_POLYGON
] = _3DPRIM_POLYGON
,
249 [PIPE_PRIM_LINES_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
250 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
251 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
252 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
253 [PIPE_PRIM_PATCHES
] = _3DPRIM_PATCHLIST_1
- 1,
256 return map
[prim
] + (prim
== PIPE_PRIM_PATCHES
? verts_per_patch
: 0);
260 translate_compare_func(enum pipe_compare_func pipe_func
)
262 static const unsigned map
[] = {
263 [PIPE_FUNC_NEVER
] = COMPAREFUNCTION_NEVER
,
264 [PIPE_FUNC_LESS
] = COMPAREFUNCTION_LESS
,
265 [PIPE_FUNC_EQUAL
] = COMPAREFUNCTION_EQUAL
,
266 [PIPE_FUNC_LEQUAL
] = COMPAREFUNCTION_LEQUAL
,
267 [PIPE_FUNC_GREATER
] = COMPAREFUNCTION_GREATER
,
268 [PIPE_FUNC_NOTEQUAL
] = COMPAREFUNCTION_NOTEQUAL
,
269 [PIPE_FUNC_GEQUAL
] = COMPAREFUNCTION_GEQUAL
,
270 [PIPE_FUNC_ALWAYS
] = COMPAREFUNCTION_ALWAYS
,
272 return map
[pipe_func
];
276 translate_shadow_func(enum pipe_compare_func pipe_func
)
278 /* Gallium specifies the result of shadow comparisons as:
280 * 1 if ref <op> texel,
285 * 0 if texel <op> ref,
288 * So we need to flip the operator and also negate.
290 static const unsigned map
[] = {
291 [PIPE_FUNC_NEVER
] = PREFILTEROPALWAYS
,
292 [PIPE_FUNC_LESS
] = PREFILTEROPLEQUAL
,
293 [PIPE_FUNC_EQUAL
] = PREFILTEROPNOTEQUAL
,
294 [PIPE_FUNC_LEQUAL
] = PREFILTEROPLESS
,
295 [PIPE_FUNC_GREATER
] = PREFILTEROPGEQUAL
,
296 [PIPE_FUNC_NOTEQUAL
] = PREFILTEROPEQUAL
,
297 [PIPE_FUNC_GEQUAL
] = PREFILTEROPGREATER
,
298 [PIPE_FUNC_ALWAYS
] = PREFILTEROPNEVER
,
300 return map
[pipe_func
];
304 translate_cull_mode(unsigned pipe_face
)
306 static const unsigned map
[4] = {
307 [PIPE_FACE_NONE
] = CULLMODE_NONE
,
308 [PIPE_FACE_FRONT
] = CULLMODE_FRONT
,
309 [PIPE_FACE_BACK
] = CULLMODE_BACK
,
310 [PIPE_FACE_FRONT_AND_BACK
] = CULLMODE_BOTH
,
312 return map
[pipe_face
];
316 translate_fill_mode(unsigned pipe_polymode
)
318 static const unsigned map
[4] = {
319 [PIPE_POLYGON_MODE_FILL
] = FILL_MODE_SOLID
,
320 [PIPE_POLYGON_MODE_LINE
] = FILL_MODE_WIREFRAME
,
321 [PIPE_POLYGON_MODE_POINT
] = FILL_MODE_POINT
,
322 [PIPE_POLYGON_MODE_FILL_RECTANGLE
] = FILL_MODE_SOLID
,
324 return map
[pipe_polymode
];
328 translate_mip_filter(enum pipe_tex_mipfilter pipe_mip
)
330 static const unsigned map
[] = {
331 [PIPE_TEX_MIPFILTER_NEAREST
] = MIPFILTER_NEAREST
,
332 [PIPE_TEX_MIPFILTER_LINEAR
] = MIPFILTER_LINEAR
,
333 [PIPE_TEX_MIPFILTER_NONE
] = MIPFILTER_NONE
,
335 return map
[pipe_mip
];
339 translate_wrap(unsigned pipe_wrap
)
341 static const unsigned map
[] = {
342 [PIPE_TEX_WRAP_REPEAT
] = TCM_WRAP
,
343 [PIPE_TEX_WRAP_CLAMP
] = TCM_HALF_BORDER
,
344 [PIPE_TEX_WRAP_CLAMP_TO_EDGE
] = TCM_CLAMP
,
345 [PIPE_TEX_WRAP_CLAMP_TO_BORDER
] = TCM_CLAMP_BORDER
,
346 [PIPE_TEX_WRAP_MIRROR_REPEAT
] = TCM_MIRROR
,
347 [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
] = TCM_MIRROR_ONCE
,
349 /* These are unsupported. */
350 [PIPE_TEX_WRAP_MIRROR_CLAMP
] = -1,
351 [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
] = -1,
353 return map
[pipe_wrap
];
356 static struct iris_address
357 ro_bo(struct iris_bo
*bo
, uint64_t offset
)
359 /* CSOs must pass NULL for bo! Otherwise it will add the BO to the
360 * validation list at CSO creation time, instead of draw time.
362 return (struct iris_address
) { .bo
= bo
, .offset
= offset
};
365 static struct iris_address
366 rw_bo(struct iris_bo
*bo
, uint64_t offset
)
368 /* CSOs must pass NULL for bo! Otherwise it will add the BO to the
369 * validation list at CSO creation time, instead of draw time.
371 return (struct iris_address
) { .bo
= bo
, .offset
= offset
, .write
= true };
375 * Allocate space for some indirect state.
377 * Return a pointer to the map (to fill it out) and a state ref (for
378 * referring to the state in GPU commands).
381 upload_state(struct u_upload_mgr
*uploader
,
382 struct iris_state_ref
*ref
,
387 u_upload_alloc(uploader
, 0, size
, alignment
, &ref
->offset
, &ref
->res
, &p
);
392 * Stream out temporary/short-lived state.
394 * This allocates space, pins the BO, and includes the BO address in the
395 * returned offset (which works because all state lives in 32-bit memory
399 stream_state(struct iris_batch
*batch
,
400 struct u_upload_mgr
*uploader
,
401 struct pipe_resource
**out_res
,
404 uint32_t *out_offset
)
408 u_upload_alloc(uploader
, 0, size
, alignment
, out_offset
, out_res
, &ptr
);
410 struct iris_bo
*bo
= iris_resource_bo(*out_res
);
411 iris_use_pinned_bo(batch
, bo
, false);
413 *out_offset
+= iris_bo_offset_from_base_address(bo
);
419 * stream_state() + memcpy.
422 emit_state(struct iris_batch
*batch
,
423 struct u_upload_mgr
*uploader
,
424 struct pipe_resource
**out_res
,
431 stream_state(batch
, uploader
, out_res
, size
, alignment
, &offset
);
434 memcpy(map
, data
, size
);
440 * Did field 'x' change between 'old_cso' and 'new_cso'?
442 * (If so, we may want to set some dirty flags.)
444 #define cso_changed(x) (!old_cso || (old_cso->x != new_cso->x))
445 #define cso_changed_memcmp(x) \
446 (!old_cso || memcmp(old_cso->x, new_cso->x, sizeof(old_cso->x)) != 0)
449 flush_for_state_base_change(struct iris_batch
*batch
)
451 /* Flush before emitting STATE_BASE_ADDRESS.
453 * This isn't documented anywhere in the PRM. However, it seems to be
454 * necessary prior to changing the surface state base adress. We've
455 * seen issues in Vulkan where we get GPU hangs when using multi-level
456 * command buffers which clear depth, reset state base address, and then
459 * Normally, in GL, we would trust the kernel to do sufficient stalls
460 * and flushes prior to executing our batch. However, it doesn't seem
461 * as if the kernel's flushing is always sufficient and we don't want to
464 * We make this an end-of-pipe sync instead of a normal flush because we
465 * do not know the current status of the GPU. On Haswell at least,
466 * having a fast-clear operation in flight at the same time as a normal
467 * rendering operation can cause hangs. Since the kernel's flushing is
468 * insufficient, we need to ensure that any rendering operations from
469 * other processes are definitely complete before we try to do our own
470 * rendering. It's a bit of a big hammer but it appears to work.
472 iris_emit_end_of_pipe_sync(batch
,
473 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
474 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
475 PIPE_CONTROL_DATA_CACHE_FLUSH
);
479 _iris_emit_lri(struct iris_batch
*batch
, uint32_t reg
, uint32_t val
)
481 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
482 lri
.RegisterOffset
= reg
;
486 #define iris_emit_lri(b, r, v) _iris_emit_lri(b, GENX(r##_num), v)
489 _iris_emit_lrr(struct iris_batch
*batch
, uint32_t dst
, uint32_t src
)
491 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_REG
), lrr
) {
492 lrr
.SourceRegisterAddress
= src
;
493 lrr
.DestinationRegisterAddress
= dst
;
498 emit_pipeline_select(struct iris_batch
*batch
, uint32_t pipeline
)
500 #if GEN_GEN >= 8 && GEN_GEN < 10
501 /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT:
503 * Software must clear the COLOR_CALC_STATE Valid field in
504 * 3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT
505 * with Pipeline Select set to GPGPU.
507 * The internal hardware docs recommend the same workaround for Gen9
510 if (pipeline
== GPGPU
)
511 iris_emit_cmd(batch
, GENX(3DSTATE_CC_STATE_POINTERS
), t
);
515 /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
516 * PIPELINE_SELECT [DevBWR+]":
520 * Software must ensure all the write caches are flushed through a
521 * stalling PIPE_CONTROL command followed by another PIPE_CONTROL
522 * command to invalidate read only caches prior to programming
523 * MI_PIPELINE_SELECT command to change the Pipeline Select Mode."
525 iris_emit_pipe_control_flush(batch
,
526 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
527 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
528 PIPE_CONTROL_DATA_CACHE_FLUSH
|
529 PIPE_CONTROL_CS_STALL
);
531 iris_emit_pipe_control_flush(batch
,
532 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
533 PIPE_CONTROL_CONST_CACHE_INVALIDATE
|
534 PIPE_CONTROL_STATE_CACHE_INVALIDATE
|
535 PIPE_CONTROL_INSTRUCTION_INVALIDATE
);
537 iris_emit_cmd(batch
, GENX(PIPELINE_SELECT
), sel
) {
541 sel
.PipelineSelection
= pipeline
;
546 init_glk_barrier_mode(struct iris_batch
*batch
, uint32_t value
)
551 * "This chicken bit works around a hardware issue with barrier
552 * logic encountered when switching between GPGPU and 3D pipelines.
553 * To workaround the issue, this mode bit should be set after a
554 * pipeline is selected."
557 iris_pack_state(GENX(SLICE_COMMON_ECO_CHICKEN1
), ®_val
, reg
) {
558 reg
.GLKBarrierMode
= value
;
559 reg
.GLKBarrierModeMask
= 1;
561 iris_emit_lri(batch
, SLICE_COMMON_ECO_CHICKEN1
, reg_val
);
566 init_state_base_address(struct iris_batch
*batch
)
568 flush_for_state_base_change(batch
);
570 /* We program most base addresses once at context initialization time.
571 * Each base address points at a 4GB memory zone, and never needs to
572 * change. See iris_bufmgr.h for a description of the memory zones.
574 * The one exception is Surface State Base Address, which needs to be
575 * updated occasionally. See iris_binder.c for the details there.
577 iris_emit_cmd(batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
579 // XXX: MOCS is stupid for this.
580 sba
.GeneralStateMemoryObjectControlState
= MOCS_WB
;
581 sba
.StatelessDataPortAccessMemoryObjectControlState
= MOCS_WB
;
582 sba
.DynamicStateMemoryObjectControlState
= MOCS_WB
;
583 sba
.IndirectObjectMemoryObjectControlState
= MOCS_WB
;
584 sba
.InstructionMemoryObjectControlState
= MOCS_WB
;
585 sba
.BindlessSurfaceStateMemoryObjectControlState
= MOCS_WB
;
588 sba
.GeneralStateBaseAddressModifyEnable
= true;
589 sba
.DynamicStateBaseAddressModifyEnable
= true;
590 sba
.IndirectObjectBaseAddressModifyEnable
= true;
591 sba
.InstructionBaseAddressModifyEnable
= true;
592 sba
.GeneralStateBufferSizeModifyEnable
= true;
593 sba
.DynamicStateBufferSizeModifyEnable
= true;
594 sba
.BindlessSurfaceStateBaseAddressModifyEnable
= true;
595 sba
.IndirectObjectBufferSizeModifyEnable
= true;
596 sba
.InstructionBuffersizeModifyEnable
= true;
598 sba
.InstructionBaseAddress
= ro_bo(NULL
, IRIS_MEMZONE_SHADER_START
);
599 sba
.DynamicStateBaseAddress
= ro_bo(NULL
, IRIS_MEMZONE_DYNAMIC_START
);
601 sba
.GeneralStateBufferSize
= 0xfffff;
602 sba
.IndirectObjectBufferSize
= 0xfffff;
603 sba
.InstructionBufferSize
= 0xfffff;
604 sba
.DynamicStateBufferSize
= 0xfffff;
609 * Upload the initial GPU state for a render context.
611 * This sets some invariant state that needs to be programmed a particular
612 * way, but we never actually change.
615 iris_init_render_context(struct iris_screen
*screen
,
616 struct iris_batch
*batch
,
617 struct iris_vtable
*vtbl
,
618 struct pipe_debug_callback
*dbg
)
620 UNUSED
const struct gen_device_info
*devinfo
= &screen
->devinfo
;
623 emit_pipeline_select(batch
, _3D
);
625 init_state_base_address(batch
);
627 // XXX: INSTPM on Gen8
628 iris_pack_state(GENX(CS_DEBUG_MODE2
), ®_val
, reg
) {
629 reg
.CONSTANT_BUFFERAddressOffsetDisable
= true;
630 reg
.CONSTANT_BUFFERAddressOffsetDisableMask
= true;
632 iris_emit_lri(batch
, CS_DEBUG_MODE2
, reg_val
);
635 iris_pack_state(GENX(CACHE_MODE_1
), ®_val
, reg
) {
636 reg
.FloatBlendOptimizationEnable
= true;
637 reg
.FloatBlendOptimizationEnableMask
= true;
638 reg
.PartialResolveDisableInVC
= true;
639 reg
.PartialResolveDisableInVCMask
= true;
641 iris_emit_lri(batch
, CACHE_MODE_1
, reg_val
);
643 if (devinfo
->is_geminilake
)
644 init_glk_barrier_mode(batch
, GLK_BARRIER_MODE_3D_HULL
);
648 iris_pack_state(GENX(SAMPLER_MODE
), ®_val
, reg
) {
649 reg
.HeaderlessMessageforPreemptableContexts
= 1;
650 reg
.HeaderlessMessageforPreemptableContextsMask
= 1;
652 iris_emit_lri(batch
, SAMPLER_MODE
, reg_val
);
657 /* 3DSTATE_DRAWING_RECTANGLE is non-pipelined, so we want to avoid
658 * changing it dynamically. We set it to the maximum size here, and
659 * instead include the render target dimensions in the viewport, so
660 * viewport extents clipping takes care of pruning stray geometry.
662 iris_emit_cmd(batch
, GENX(3DSTATE_DRAWING_RECTANGLE
), rect
) {
663 rect
.ClippedDrawingRectangleXMax
= UINT16_MAX
;
664 rect
.ClippedDrawingRectangleYMax
= UINT16_MAX
;
667 /* Set the initial MSAA sample positions. */
668 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLE_PATTERN
), pat
) {
669 GEN_SAMPLE_POS_1X(pat
._1xSample
);
670 GEN_SAMPLE_POS_2X(pat
._2xSample
);
671 GEN_SAMPLE_POS_4X(pat
._4xSample
);
672 GEN_SAMPLE_POS_8X(pat
._8xSample
);
673 GEN_SAMPLE_POS_16X(pat
._16xSample
);
676 /* Use the legacy AA line coverage computation. */
677 iris_emit_cmd(batch
, GENX(3DSTATE_AA_LINE_PARAMETERS
), foo
);
679 /* Disable chromakeying (it's for media) */
680 iris_emit_cmd(batch
, GENX(3DSTATE_WM_CHROMAKEY
), foo
);
682 /* We want regular rendering, not special HiZ operations. */
683 iris_emit_cmd(batch
, GENX(3DSTATE_WM_HZ_OP
), foo
);
685 /* No polygon stippling offsets are necessary. */
686 // XXX: may need to set an offset for origin-UL framebuffers
687 iris_emit_cmd(batch
, GENX(3DSTATE_POLY_STIPPLE_OFFSET
), foo
);
689 /* Set a static partitioning of the push constant area. */
690 // XXX: this may be a bad idea...could starve the push ringbuffers...
691 for (int i
= 0; i
<= MESA_SHADER_FRAGMENT
; i
++) {
692 iris_emit_cmd(batch
, GENX(3DSTATE_PUSH_CONSTANT_ALLOC_VS
), alloc
) {
693 alloc
._3DCommandSubOpcode
= 18 + i
;
694 alloc
.ConstantBufferOffset
= 6 * i
;
695 alloc
.ConstantBufferSize
= i
== MESA_SHADER_FRAGMENT
? 8 : 6;
701 iris_init_compute_context(struct iris_screen
*screen
,
702 struct iris_batch
*batch
,
703 struct iris_vtable
*vtbl
,
704 struct pipe_debug_callback
*dbg
)
706 UNUSED
const struct gen_device_info
*devinfo
= &screen
->devinfo
;
708 emit_pipeline_select(batch
, GPGPU
);
710 const bool has_slm
= true;
711 const bool wants_dc_cache
= true;
713 const struct gen_l3_weights w
=
714 gen_get_default_l3_weights(devinfo
, wants_dc_cache
, has_slm
);
715 const struct gen_l3_config
*cfg
= gen_get_l3_config(devinfo
, w
);
718 iris_pack_state(GENX(L3CNTLREG
), ®_val
, reg
) {
719 reg
.SLMEnable
= has_slm
;
721 /* WA_1406697149: Bit 9 "Error Detection Behavior Control" must be set
722 * in L3CNTLREG register. The default setting of the bit is not the
723 * desirable behavior.
725 reg
.ErrorDetectionBehaviorControl
= true;
727 reg
.URBAllocation
= cfg
->n
[GEN_L3P_URB
];
728 reg
.ROAllocation
= cfg
->n
[GEN_L3P_RO
];
729 reg
.DCAllocation
= cfg
->n
[GEN_L3P_DC
];
730 reg
.AllAllocation
= cfg
->n
[GEN_L3P_ALL
];
732 iris_emit_lri(batch
, L3CNTLREG
, reg_val
);
734 init_state_base_address(batch
);
737 if (devinfo
->is_geminilake
)
738 init_glk_barrier_mode(batch
, GLK_BARRIER_MODE_GPGPU
);
742 struct iris_vertex_buffer_state
{
743 /** The VERTEX_BUFFER_STATE hardware structure. */
744 uint32_t state
[GENX(VERTEX_BUFFER_STATE_length
)];
746 /** The resource to source vertex data from. */
747 struct pipe_resource
*resource
;
750 struct iris_depth_buffer_state
{
751 /* Depth/HiZ/Stencil related hardware packets. */
752 uint32_t packets
[GENX(3DSTATE_DEPTH_BUFFER_length
) +
753 GENX(3DSTATE_STENCIL_BUFFER_length
) +
754 GENX(3DSTATE_HIER_DEPTH_BUFFER_length
) +
755 GENX(3DSTATE_CLEAR_PARAMS_length
)];
759 * Generation-specific context state (ice->state.genx->...).
761 * Most state can go in iris_context directly, but these encode hardware
762 * packets which vary by generation.
764 struct iris_genx_state
{
765 struct iris_vertex_buffer_state vertex_buffers
[33];
767 /** The number of bound vertex buffers. */
768 uint64_t bound_vertex_buffers
;
770 struct iris_depth_buffer_state depth_buffer
;
772 uint32_t so_buffers
[4 * GENX(3DSTATE_SO_BUFFER_length
)];
773 uint32_t streamout
[4 * GENX(3DSTATE_STREAMOUT_length
)];
777 * The pipe->set_blend_color() driver hook.
779 * This corresponds to our COLOR_CALC_STATE.
782 iris_set_blend_color(struct pipe_context
*ctx
,
783 const struct pipe_blend_color
*state
)
785 struct iris_context
*ice
= (struct iris_context
*) ctx
;
787 /* Our COLOR_CALC_STATE is exactly pipe_blend_color, so just memcpy */
788 memcpy(&ice
->state
.blend_color
, state
, sizeof(struct pipe_blend_color
));
789 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
793 * Gallium CSO for blend state (see pipe_blend_state).
795 struct iris_blend_state
{
796 /** Partial 3DSTATE_PS_BLEND */
797 uint32_t ps_blend
[GENX(3DSTATE_PS_BLEND_length
)];
799 /** Partial BLEND_STATE */
800 uint32_t blend_state
[GENX(BLEND_STATE_length
) +
801 BRW_MAX_DRAW_BUFFERS
* GENX(BLEND_STATE_ENTRY_length
)];
803 bool alpha_to_coverage
; /* for shader key */
806 static enum pipe_blendfactor
807 fix_blendfactor(enum pipe_blendfactor f
, bool alpha_to_one
)
810 if (f
== PIPE_BLENDFACTOR_SRC1_ALPHA
)
811 return PIPE_BLENDFACTOR_ONE
;
813 if (f
== PIPE_BLENDFACTOR_INV_SRC1_ALPHA
)
814 return PIPE_BLENDFACTOR_ZERO
;
821 * The pipe->create_blend_state() driver hook.
823 * Translates a pipe_blend_state into iris_blend_state.
826 iris_create_blend_state(struct pipe_context
*ctx
,
827 const struct pipe_blend_state
*state
)
829 struct iris_blend_state
*cso
= malloc(sizeof(struct iris_blend_state
));
830 uint32_t *blend_entry
= cso
->blend_state
+ GENX(BLEND_STATE_length
);
832 cso
->alpha_to_coverage
= state
->alpha_to_coverage
;
834 bool indep_alpha_blend
= false;
836 for (int i
= 0; i
< BRW_MAX_DRAW_BUFFERS
; i
++) {
837 const struct pipe_rt_blend_state
*rt
=
838 &state
->rt
[state
->independent_blend_enable
? i
: 0];
840 enum pipe_blendfactor src_rgb
=
841 fix_blendfactor(rt
->rgb_src_factor
, state
->alpha_to_one
);
842 enum pipe_blendfactor src_alpha
=
843 fix_blendfactor(rt
->alpha_src_factor
, state
->alpha_to_one
);
844 enum pipe_blendfactor dst_rgb
=
845 fix_blendfactor(rt
->rgb_dst_factor
, state
->alpha_to_one
);
846 enum pipe_blendfactor dst_alpha
=
847 fix_blendfactor(rt
->alpha_dst_factor
, state
->alpha_to_one
);
849 if (rt
->rgb_func
!= rt
->alpha_func
||
850 src_rgb
!= src_alpha
|| dst_rgb
!= dst_alpha
)
851 indep_alpha_blend
= true;
853 iris_pack_state(GENX(BLEND_STATE_ENTRY
), blend_entry
, be
) {
854 be
.LogicOpEnable
= state
->logicop_enable
;
855 be
.LogicOpFunction
= state
->logicop_func
;
857 be
.PreBlendSourceOnlyClampEnable
= false;
858 be
.ColorClampRange
= COLORCLAMP_RTFORMAT
;
859 be
.PreBlendColorClampEnable
= true;
860 be
.PostBlendColorClampEnable
= true;
862 be
.ColorBufferBlendEnable
= rt
->blend_enable
;
864 be
.ColorBlendFunction
= rt
->rgb_func
;
865 be
.AlphaBlendFunction
= rt
->alpha_func
;
866 be
.SourceBlendFactor
= src_rgb
;
867 be
.SourceAlphaBlendFactor
= src_alpha
;
868 be
.DestinationBlendFactor
= dst_rgb
;
869 be
.DestinationAlphaBlendFactor
= dst_alpha
;
871 be
.WriteDisableRed
= !(rt
->colormask
& PIPE_MASK_R
);
872 be
.WriteDisableGreen
= !(rt
->colormask
& PIPE_MASK_G
);
873 be
.WriteDisableBlue
= !(rt
->colormask
& PIPE_MASK_B
);
874 be
.WriteDisableAlpha
= !(rt
->colormask
& PIPE_MASK_A
);
876 blend_entry
+= GENX(BLEND_STATE_ENTRY_length
);
879 iris_pack_command(GENX(3DSTATE_PS_BLEND
), cso
->ps_blend
, pb
) {
880 /* pb.HasWriteableRT is filled in at draw time. */
881 /* pb.AlphaTestEnable is filled in at draw time. */
882 pb
.AlphaToCoverageEnable
= state
->alpha_to_coverage
;
883 pb
.IndependentAlphaBlendEnable
= indep_alpha_blend
;
885 pb
.ColorBufferBlendEnable
= state
->rt
[0].blend_enable
;
887 pb
.SourceBlendFactor
=
888 fix_blendfactor(state
->rt
[0].rgb_src_factor
, state
->alpha_to_one
);
889 pb
.SourceAlphaBlendFactor
=
890 fix_blendfactor(state
->rt
[0].alpha_src_factor
, state
->alpha_to_one
);
891 pb
.DestinationBlendFactor
=
892 fix_blendfactor(state
->rt
[0].rgb_dst_factor
, state
->alpha_to_one
);
893 pb
.DestinationAlphaBlendFactor
=
894 fix_blendfactor(state
->rt
[0].alpha_dst_factor
, state
->alpha_to_one
);
897 iris_pack_state(GENX(BLEND_STATE
), cso
->blend_state
, bs
) {
898 bs
.AlphaToCoverageEnable
= state
->alpha_to_coverage
;
899 bs
.IndependentAlphaBlendEnable
= indep_alpha_blend
;
900 bs
.AlphaToOneEnable
= state
->alpha_to_one
;
901 bs
.AlphaToCoverageDitherEnable
= state
->alpha_to_coverage
;
902 bs
.ColorDitherEnable
= state
->dither
;
903 /* bl.AlphaTestEnable and bs.AlphaTestFunction are filled in later. */
911 * The pipe->bind_blend_state() driver hook.
913 * Bind a blending CSO and flag related dirty bits.
916 iris_bind_blend_state(struct pipe_context
*ctx
, void *state
)
918 struct iris_context
*ice
= (struct iris_context
*) ctx
;
919 ice
->state
.cso_blend
= state
;
920 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
;
921 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
922 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_BLEND
];
926 * Gallium CSO for depth, stencil, and alpha testing state.
928 struct iris_depth_stencil_alpha_state
{
929 /** Partial 3DSTATE_WM_DEPTH_STENCIL. */
930 uint32_t wmds
[GENX(3DSTATE_WM_DEPTH_STENCIL_length
)];
932 /** Outbound to BLEND_STATE, 3DSTATE_PS_BLEND, COLOR_CALC_STATE. */
933 struct pipe_alpha_state alpha
;
935 /** Outbound to resolve and cache set tracking. */
936 bool depth_writes_enabled
;
937 bool stencil_writes_enabled
;
941 * The pipe->create_depth_stencil_alpha_state() driver hook.
943 * We encode most of 3DSTATE_WM_DEPTH_STENCIL, and just save off the alpha
944 * testing state since we need pieces of it in a variety of places.
947 iris_create_zsa_state(struct pipe_context
*ctx
,
948 const struct pipe_depth_stencil_alpha_state
*state
)
950 struct iris_depth_stencil_alpha_state
*cso
=
951 malloc(sizeof(struct iris_depth_stencil_alpha_state
));
953 bool two_sided_stencil
= state
->stencil
[1].enabled
;
955 cso
->alpha
= state
->alpha
;
956 cso
->depth_writes_enabled
= state
->depth
.writemask
;
957 cso
->stencil_writes_enabled
=
958 state
->stencil
[0].writemask
!= 0 ||
959 (two_sided_stencil
&& state
->stencil
[1].writemask
!= 1);
961 /* The state tracker needs to optimize away EQUAL writes for us. */
962 assert(!(state
->depth
.func
== PIPE_FUNC_EQUAL
&& state
->depth
.writemask
));
964 iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL
), cso
->wmds
, wmds
) {
965 wmds
.StencilFailOp
= state
->stencil
[0].fail_op
;
966 wmds
.StencilPassDepthFailOp
= state
->stencil
[0].zfail_op
;
967 wmds
.StencilPassDepthPassOp
= state
->stencil
[0].zpass_op
;
968 wmds
.StencilTestFunction
=
969 translate_compare_func(state
->stencil
[0].func
);
970 wmds
.BackfaceStencilFailOp
= state
->stencil
[1].fail_op
;
971 wmds
.BackfaceStencilPassDepthFailOp
= state
->stencil
[1].zfail_op
;
972 wmds
.BackfaceStencilPassDepthPassOp
= state
->stencil
[1].zpass_op
;
973 wmds
.BackfaceStencilTestFunction
=
974 translate_compare_func(state
->stencil
[1].func
);
975 wmds
.DepthTestFunction
= translate_compare_func(state
->depth
.func
);
976 wmds
.DoubleSidedStencilEnable
= two_sided_stencil
;
977 wmds
.StencilTestEnable
= state
->stencil
[0].enabled
;
978 wmds
.StencilBufferWriteEnable
=
979 state
->stencil
[0].writemask
!= 0 ||
980 (two_sided_stencil
&& state
->stencil
[1].writemask
!= 0);
981 wmds
.DepthTestEnable
= state
->depth
.enabled
;
982 wmds
.DepthBufferWriteEnable
= state
->depth
.writemask
;
983 wmds
.StencilTestMask
= state
->stencil
[0].valuemask
;
984 wmds
.StencilWriteMask
= state
->stencil
[0].writemask
;
985 wmds
.BackfaceStencilTestMask
= state
->stencil
[1].valuemask
;
986 wmds
.BackfaceStencilWriteMask
= state
->stencil
[1].writemask
;
987 /* wmds.[Backface]StencilReferenceValue are merged later */
994 * The pipe->bind_depth_stencil_alpha_state() driver hook.
996 * Bind a depth/stencil/alpha CSO and flag related dirty bits.
999 iris_bind_zsa_state(struct pipe_context
*ctx
, void *state
)
1001 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1002 struct iris_depth_stencil_alpha_state
*old_cso
= ice
->state
.cso_zsa
;
1003 struct iris_depth_stencil_alpha_state
*new_cso
= state
;
1006 if (cso_changed(alpha
.ref_value
))
1007 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
1009 if (cso_changed(alpha
.enabled
))
1010 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
| IRIS_DIRTY_BLEND_STATE
;
1012 if (cso_changed(alpha
.func
))
1013 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
1015 ice
->state
.depth_writes_enabled
= new_cso
->depth_writes_enabled
;
1016 ice
->state
.stencil_writes_enabled
= new_cso
->stencil_writes_enabled
;
1019 ice
->state
.cso_zsa
= new_cso
;
1020 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
1021 ice
->state
.dirty
|= IRIS_DIRTY_WM_DEPTH_STENCIL
;
1022 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_DEPTH_STENCIL_ALPHA
];
1026 * Gallium CSO for rasterizer state.
1028 struct iris_rasterizer_state
{
1029 uint32_t sf
[GENX(3DSTATE_SF_length
)];
1030 uint32_t clip
[GENX(3DSTATE_CLIP_length
)];
1031 uint32_t raster
[GENX(3DSTATE_RASTER_length
)];
1032 uint32_t wm
[GENX(3DSTATE_WM_length
)];
1033 uint32_t line_stipple
[GENX(3DSTATE_LINE_STIPPLE_length
)];
1035 uint8_t num_clip_plane_consts
;
1036 bool clip_halfz
; /* for CC_VIEWPORT */
1037 bool depth_clip_near
; /* for CC_VIEWPORT */
1038 bool depth_clip_far
; /* for CC_VIEWPORT */
1039 bool flatshade
; /* for shader state */
1040 bool flatshade_first
; /* for stream output */
1041 bool clamp_fragment_color
; /* for shader state */
1042 bool light_twoside
; /* for shader state */
1043 bool rasterizer_discard
; /* for 3DSTATE_STREAMOUT and 3DSTATE_CLIP */
1044 bool half_pixel_center
; /* for 3DSTATE_MULTISAMPLE */
1045 bool line_stipple_enable
;
1046 bool poly_stipple_enable
;
1048 bool force_persample_interp
;
1049 enum pipe_sprite_coord_mode sprite_coord_mode
; /* PIPE_SPRITE_* */
1050 uint16_t sprite_coord_enable
;
1054 get_line_width(const struct pipe_rasterizer_state
*state
)
1056 float line_width
= state
->line_width
;
1058 /* From the OpenGL 4.4 spec:
1060 * "The actual width of non-antialiased lines is determined by rounding
1061 * the supplied width to the nearest integer, then clamping it to the
1062 * implementation-dependent maximum non-antialiased line width."
1064 if (!state
->multisample
&& !state
->line_smooth
)
1065 line_width
= roundf(state
->line_width
);
1067 if (!state
->multisample
&& state
->line_smooth
&& line_width
< 1.5f
) {
1068 /* For 1 pixel line thickness or less, the general anti-aliasing
1069 * algorithm gives up, and a garbage line is generated. Setting a
1070 * Line Width of 0.0 specifies the rasterization of the "thinnest"
1071 * (one-pixel-wide), non-antialiased lines.
1073 * Lines rendered with zero Line Width are rasterized using the
1074 * "Grid Intersection Quantization" rules as specified by the
1075 * "Zero-Width (Cosmetic) Line Rasterization" section of the docs.
1084 * The pipe->create_rasterizer_state() driver hook.
1087 iris_create_rasterizer_state(struct pipe_context
*ctx
,
1088 const struct pipe_rasterizer_state
*state
)
1090 struct iris_rasterizer_state
*cso
=
1091 malloc(sizeof(struct iris_rasterizer_state
));
1094 point_quad_rasterization
-> SBE
?
1101 offset_units_unscaled
- cap
not exposed
1105 // XXX: it may make more sense just to store the pipe_rasterizer_state,
1106 // we're copying a lot of booleans here. But we don't need all of them...
1108 cso
->multisample
= state
->multisample
;
1109 cso
->force_persample_interp
= state
->force_persample_interp
;
1110 cso
->clip_halfz
= state
->clip_halfz
;
1111 cso
->depth_clip_near
= state
->depth_clip_near
;
1112 cso
->depth_clip_far
= state
->depth_clip_far
;
1113 cso
->flatshade
= state
->flatshade
;
1114 cso
->flatshade_first
= state
->flatshade_first
;
1115 cso
->clamp_fragment_color
= state
->clamp_fragment_color
;
1116 cso
->light_twoside
= state
->light_twoside
;
1117 cso
->rasterizer_discard
= state
->rasterizer_discard
;
1118 cso
->half_pixel_center
= state
->half_pixel_center
;
1119 cso
->sprite_coord_mode
= state
->sprite_coord_mode
;
1120 cso
->sprite_coord_enable
= state
->sprite_coord_enable
;
1121 cso
->line_stipple_enable
= state
->line_stipple_enable
;
1122 cso
->poly_stipple_enable
= state
->poly_stipple_enable
;
1124 if (state
->clip_plane_enable
!= 0)
1125 cso
->num_clip_plane_consts
= util_logbase2(state
->clip_plane_enable
) + 1;
1127 cso
->num_clip_plane_consts
= 0;
1129 float line_width
= get_line_width(state
);
1131 iris_pack_command(GENX(3DSTATE_SF
), cso
->sf
, sf
) {
1132 sf
.StatisticsEnable
= true;
1133 sf
.ViewportTransformEnable
= true;
1134 sf
.AALineDistanceMode
= AALINEDISTANCE_TRUE
;
1135 sf
.LineEndCapAntialiasingRegionWidth
=
1136 state
->line_smooth
? _10pixels
: _05pixels
;
1137 sf
.LastPixelEnable
= state
->line_last_pixel
;
1138 sf
.LineWidth
= line_width
;
1139 sf
.SmoothPointEnable
= state
->point_smooth
|| state
->multisample
;
1140 sf
.PointWidthSource
= state
->point_size_per_vertex
? Vertex
: State
;
1141 sf
.PointWidth
= state
->point_size
;
1143 if (state
->flatshade_first
) {
1144 sf
.TriangleFanProvokingVertexSelect
= 1;
1146 sf
.TriangleStripListProvokingVertexSelect
= 2;
1147 sf
.TriangleFanProvokingVertexSelect
= 2;
1148 sf
.LineStripListProvokingVertexSelect
= 1;
1152 iris_pack_command(GENX(3DSTATE_RASTER
), cso
->raster
, rr
) {
1153 rr
.FrontWinding
= state
->front_ccw
? CounterClockwise
: Clockwise
;
1154 rr
.CullMode
= translate_cull_mode(state
->cull_face
);
1155 rr
.FrontFaceFillMode
= translate_fill_mode(state
->fill_front
);
1156 rr
.BackFaceFillMode
= translate_fill_mode(state
->fill_back
);
1157 rr
.DXMultisampleRasterizationEnable
= state
->multisample
;
1158 rr
.GlobalDepthOffsetEnableSolid
= state
->offset_tri
;
1159 rr
.GlobalDepthOffsetEnableWireframe
= state
->offset_line
;
1160 rr
.GlobalDepthOffsetEnablePoint
= state
->offset_point
;
1161 rr
.GlobalDepthOffsetConstant
= state
->offset_units
* 2;
1162 rr
.GlobalDepthOffsetScale
= state
->offset_scale
;
1163 rr
.GlobalDepthOffsetClamp
= state
->offset_clamp
;
1164 rr
.SmoothPointEnable
= state
->point_smooth
|| state
->multisample
;
1165 rr
.AntialiasingEnable
= state
->line_smooth
;
1166 rr
.ScissorRectangleEnable
= state
->scissor
;
1167 rr
.ViewportZNearClipTestEnable
= state
->depth_clip_near
;
1168 rr
.ViewportZFarClipTestEnable
= state
->depth_clip_far
;
1169 //rr.ConservativeRasterizationEnable = not yet supported by Gallium...
1172 iris_pack_command(GENX(3DSTATE_CLIP
), cso
->clip
, cl
) {
1173 /* cl.NonPerspectiveBarycentricEnable is filled in at draw time from
1174 * the FS program; cl.ForceZeroRTAIndexEnable is filled in from the FB.
1176 cl
.EarlyCullEnable
= true;
1177 cl
.UserClipDistanceClipTestEnableBitmask
= state
->clip_plane_enable
;
1178 cl
.ForceUserClipDistanceClipTestEnableBitmask
= true;
1179 cl
.APIMode
= state
->clip_halfz
? APIMODE_D3D
: APIMODE_OGL
;
1180 cl
.GuardbandClipTestEnable
= true;
1181 cl
.ClipEnable
= true;
1182 cl
.ViewportXYClipTestEnable
= state
->point_tri_clip
;
1183 cl
.MinimumPointWidth
= 0.125;
1184 cl
.MaximumPointWidth
= 255.875;
1186 if (state
->flatshade_first
) {
1187 cl
.TriangleFanProvokingVertexSelect
= 1;
1189 cl
.TriangleStripListProvokingVertexSelect
= 2;
1190 cl
.TriangleFanProvokingVertexSelect
= 2;
1191 cl
.LineStripListProvokingVertexSelect
= 1;
1195 iris_pack_command(GENX(3DSTATE_WM
), cso
->wm
, wm
) {
1196 /* wm.BarycentricInterpolationMode and wm.EarlyDepthStencilControl are
1197 * filled in at draw time from the FS program.
1199 wm
.LineAntialiasingRegionWidth
= _10pixels
;
1200 wm
.LineEndCapAntialiasingRegionWidth
= _05pixels
;
1201 wm
.PointRasterizationRule
= RASTRULE_UPPER_RIGHT
;
1202 wm
.LineStippleEnable
= state
->line_stipple_enable
;
1203 wm
.PolygonStippleEnable
= state
->poly_stipple_enable
;
1206 /* Remap from 0..255 back to 1..256 */
1207 const unsigned line_stipple_factor
= state
->line_stipple_factor
+ 1;
1209 iris_pack_command(GENX(3DSTATE_LINE_STIPPLE
), cso
->line_stipple
, line
) {
1210 line
.LineStipplePattern
= state
->line_stipple_pattern
;
1211 line
.LineStippleInverseRepeatCount
= 1.0f
/ line_stipple_factor
;
1212 line
.LineStippleRepeatCount
= line_stipple_factor
;
1219 * The pipe->bind_rasterizer_state() driver hook.
1221 * Bind a rasterizer CSO and flag related dirty bits.
1224 iris_bind_rasterizer_state(struct pipe_context
*ctx
, void *state
)
1226 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1227 struct iris_rasterizer_state
*old_cso
= ice
->state
.cso_rast
;
1228 struct iris_rasterizer_state
*new_cso
= state
;
1231 /* Try to avoid re-emitting 3DSTATE_LINE_STIPPLE, it's non-pipelined */
1232 if (cso_changed_memcmp(line_stipple
))
1233 ice
->state
.dirty
|= IRIS_DIRTY_LINE_STIPPLE
;
1235 if (cso_changed(half_pixel_center
))
1236 ice
->state
.dirty
|= IRIS_DIRTY_MULTISAMPLE
;
1238 if (cso_changed(line_stipple_enable
) || cso_changed(poly_stipple_enable
))
1239 ice
->state
.dirty
|= IRIS_DIRTY_WM
;
1241 if (cso_changed(rasterizer_discard
))
1242 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
| IRIS_DIRTY_CLIP
;
1244 if (cso_changed(flatshade_first
))
1245 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
;
1247 if (cso_changed(depth_clip_near
) || cso_changed(depth_clip_far
) ||
1248 cso_changed(clip_halfz
))
1249 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
1251 if (cso_changed(sprite_coord_enable
) ||
1252 cso_changed(sprite_coord_mode
) ||
1253 cso_changed(light_twoside
))
1254 ice
->state
.dirty
|= IRIS_DIRTY_SBE
;
1257 ice
->state
.cso_rast
= new_cso
;
1258 ice
->state
.dirty
|= IRIS_DIRTY_RASTER
;
1259 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
1260 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_RASTERIZER
];
1264 * Return true if the given wrap mode requires the border color to exist.
1266 * (We can skip uploading it if the sampler isn't going to use it.)
1269 wrap_mode_needs_border_color(unsigned wrap_mode
)
1271 return wrap_mode
== TCM_CLAMP_BORDER
|| wrap_mode
== TCM_HALF_BORDER
;
1275 * Gallium CSO for sampler state.
1277 struct iris_sampler_state
{
1278 union pipe_color_union border_color
;
1279 bool needs_border_color
;
1281 uint32_t sampler_state
[GENX(SAMPLER_STATE_length
)];
1285 * The pipe->create_sampler_state() driver hook.
1287 * We fill out SAMPLER_STATE (except for the border color pointer), and
1288 * store that on the CPU. It doesn't make sense to upload it to a GPU
1289 * buffer object yet, because 3DSTATE_SAMPLER_STATE_POINTERS requires
1290 * all bound sampler states to be in contiguous memor.
1293 iris_create_sampler_state(struct pipe_context
*ctx
,
1294 const struct pipe_sampler_state
*state
)
1296 struct iris_sampler_state
*cso
= CALLOC_STRUCT(iris_sampler_state
);
1301 STATIC_ASSERT(PIPE_TEX_FILTER_NEAREST
== MAPFILTER_NEAREST
);
1302 STATIC_ASSERT(PIPE_TEX_FILTER_LINEAR
== MAPFILTER_LINEAR
);
1304 unsigned wrap_s
= translate_wrap(state
->wrap_s
);
1305 unsigned wrap_t
= translate_wrap(state
->wrap_t
);
1306 unsigned wrap_r
= translate_wrap(state
->wrap_r
);
1308 memcpy(&cso
->border_color
, &state
->border_color
, sizeof(cso
->border_color
));
1310 cso
->needs_border_color
= wrap_mode_needs_border_color(wrap_s
) ||
1311 wrap_mode_needs_border_color(wrap_t
) ||
1312 wrap_mode_needs_border_color(wrap_r
);
1314 float min_lod
= state
->min_lod
;
1315 unsigned mag_img_filter
= state
->mag_img_filter
;
1317 // XXX: explain this code ported from ilo...I don't get it at all...
1318 if (state
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
&&
1319 state
->min_lod
> 0.0f
) {
1321 mag_img_filter
= state
->min_img_filter
;
1324 iris_pack_state(GENX(SAMPLER_STATE
), cso
->sampler_state
, samp
) {
1325 samp
.TCXAddressControlMode
= wrap_s
;
1326 samp
.TCYAddressControlMode
= wrap_t
;
1327 samp
.TCZAddressControlMode
= wrap_r
;
1328 samp
.CubeSurfaceControlMode
= state
->seamless_cube_map
;
1329 samp
.NonnormalizedCoordinateEnable
= !state
->normalized_coords
;
1330 samp
.MinModeFilter
= state
->min_img_filter
;
1331 samp
.MagModeFilter
= mag_img_filter
;
1332 samp
.MipModeFilter
= translate_mip_filter(state
->min_mip_filter
);
1333 samp
.MaximumAnisotropy
= RATIO21
;
1335 if (state
->max_anisotropy
>= 2) {
1336 if (state
->min_img_filter
== PIPE_TEX_FILTER_LINEAR
) {
1337 samp
.MinModeFilter
= MAPFILTER_ANISOTROPIC
;
1338 samp
.AnisotropicAlgorithm
= EWAApproximation
;
1341 if (state
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
)
1342 samp
.MagModeFilter
= MAPFILTER_ANISOTROPIC
;
1344 samp
.MaximumAnisotropy
=
1345 MIN2((state
->max_anisotropy
- 2) / 2, RATIO161
);
1348 /* Set address rounding bits if not using nearest filtering. */
1349 if (state
->min_img_filter
!= PIPE_TEX_FILTER_NEAREST
) {
1350 samp
.UAddressMinFilterRoundingEnable
= true;
1351 samp
.VAddressMinFilterRoundingEnable
= true;
1352 samp
.RAddressMinFilterRoundingEnable
= true;
1355 if (state
->mag_img_filter
!= PIPE_TEX_FILTER_NEAREST
) {
1356 samp
.UAddressMagFilterRoundingEnable
= true;
1357 samp
.VAddressMagFilterRoundingEnable
= true;
1358 samp
.RAddressMagFilterRoundingEnable
= true;
1361 if (state
->compare_mode
== PIPE_TEX_COMPARE_R_TO_TEXTURE
)
1362 samp
.ShadowFunction
= translate_shadow_func(state
->compare_func
);
1364 const float hw_max_lod
= GEN_GEN
>= 7 ? 14 : 13;
1366 samp
.LODPreClampMode
= CLAMP_MODE_OGL
;
1367 samp
.MinLOD
= CLAMP(min_lod
, 0, hw_max_lod
);
1368 samp
.MaxLOD
= CLAMP(state
->max_lod
, 0, hw_max_lod
);
1369 samp
.TextureLODBias
= CLAMP(state
->lod_bias
, -16, 15);
1371 /* .BorderColorPointer is filled in by iris_bind_sampler_states. */
1378 * The pipe->bind_sampler_states() driver hook.
1380 * Now that we know all the sampler states, we upload them all into a
1381 * contiguous area of GPU memory, for 3DSTATE_SAMPLER_STATE_POINTERS_*.
1382 * We also fill out the border color state pointers at this point.
1384 * We could defer this work to draw time, but we assume that binding
1385 * will be less frequent than drawing.
1387 // XXX: this may be a bad idea, need to make sure that st/mesa calls us
1388 // XXX: with the complete set of shaders. If it makes multiple calls to
1389 // XXX: things one at a time, we could waste a lot of time assembling things.
1390 // XXX: it doesn't even BUY us anything to do it here, because we only flag
1391 // XXX: IRIS_DIRTY_SAMPLER_STATE when this is called...
1393 iris_bind_sampler_states(struct pipe_context
*ctx
,
1394 enum pipe_shader_type p_stage
,
1395 unsigned start
, unsigned count
,
1398 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1399 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1400 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1402 assert(start
+ count
<= IRIS_MAX_TEXTURE_SAMPLERS
);
1404 for (int i
= 0; i
< count
; i
++) {
1405 shs
->samplers
[start
+ i
] = states
[i
];
1408 /* Assemble the SAMPLER_STATEs into a contiguous table that lives
1409 * in the dynamic state memory zone, so we can point to it via the
1410 * 3DSTATE_SAMPLER_STATE_POINTERS_* commands.
1413 upload_state(ice
->state
.dynamic_uploader
, &shs
->sampler_table
,
1414 count
* 4 * GENX(SAMPLER_STATE_length
), 32);
1418 struct pipe_resource
*res
= shs
->sampler_table
.res
;
1419 shs
->sampler_table
.offset
+=
1420 iris_bo_offset_from_base_address(iris_resource_bo(res
));
1422 /* Make sure all land in the same BO */
1423 iris_border_color_pool_reserve(ice
, IRIS_MAX_TEXTURE_SAMPLERS
);
1425 for (int i
= 0; i
< count
; i
++) {
1426 struct iris_sampler_state
*state
= shs
->samplers
[i
];
1429 memset(map
, 0, 4 * GENX(SAMPLER_STATE_length
));
1430 } else if (!state
->needs_border_color
) {
1431 memcpy(map
, state
->sampler_state
, 4 * GENX(SAMPLER_STATE_length
));
1433 ice
->state
.need_border_colors
= true;
1435 /* Stream out the border color and merge the pointer. */
1437 iris_upload_border_color(ice
, &state
->border_color
);
1439 uint32_t dynamic
[GENX(SAMPLER_STATE_length
)];
1440 iris_pack_state(GENX(SAMPLER_STATE
), dynamic
, dyns
) {
1441 dyns
.BorderColorPointer
= offset
;
1444 for (uint32_t j
= 0; j
< GENX(SAMPLER_STATE_length
); j
++)
1445 map
[j
] = state
->sampler_state
[j
] | dynamic
[j
];
1448 map
+= GENX(SAMPLER_STATE_length
);
1451 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
;
1454 static enum isl_channel_select
1455 fmt_swizzle(const struct iris_format_info
*fmt
, enum pipe_swizzle swz
)
1458 case PIPE_SWIZZLE_X
: return fmt
->swizzle
.r
;
1459 case PIPE_SWIZZLE_Y
: return fmt
->swizzle
.g
;
1460 case PIPE_SWIZZLE_Z
: return fmt
->swizzle
.b
;
1461 case PIPE_SWIZZLE_W
: return fmt
->swizzle
.a
;
1462 case PIPE_SWIZZLE_1
: return SCS_ONE
;
1463 case PIPE_SWIZZLE_0
: return SCS_ZERO
;
1464 default: unreachable("invalid swizzle");
1469 fill_buffer_surface_state(struct isl_device
*isl_dev
,
1472 enum isl_format format
,
1476 const struct isl_format_layout
*fmtl
= isl_format_get_layout(format
);
1477 const unsigned cpp
= fmtl
->bpb
/ 8;
1479 /* The ARB_texture_buffer_specification says:
1481 * "The number of texels in the buffer texture's texel array is given by
1483 * floor(<buffer_size> / (<components> * sizeof(<base_type>)),
1485 * where <buffer_size> is the size of the buffer object, in basic
1486 * machine units and <components> and <base_type> are the element count
1487 * and base data type for elements, as specified in Table X.1. The
1488 * number of texels in the texel array is then clamped to the
1489 * implementation-dependent limit MAX_TEXTURE_BUFFER_SIZE_ARB."
1491 * We need to clamp the size in bytes to MAX_TEXTURE_BUFFER_SIZE * stride,
1492 * so that when ISL divides by stride to obtain the number of texels, that
1493 * texel count is clamped to MAX_TEXTURE_BUFFER_SIZE.
1495 unsigned final_size
=
1496 MIN3(size
, bo
->size
- offset
, IRIS_MAX_TEXTURE_BUFFER_SIZE
* cpp
);
1498 isl_buffer_fill_state(isl_dev
, map
,
1499 .address
= bo
->gtt_offset
+ offset
,
1500 .size_B
= final_size
,
1507 * The pipe->create_sampler_view() driver hook.
1509 static struct pipe_sampler_view
*
1510 iris_create_sampler_view(struct pipe_context
*ctx
,
1511 struct pipe_resource
*tex
,
1512 const struct pipe_sampler_view
*tmpl
)
1514 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1515 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1516 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1517 struct iris_sampler_view
*isv
= calloc(1, sizeof(struct iris_sampler_view
));
1522 /* initialize base object */
1524 isv
->base
.context
= ctx
;
1525 isv
->base
.texture
= NULL
;
1526 pipe_reference_init(&isv
->base
.reference
, 1);
1527 pipe_resource_reference(&isv
->base
.texture
, tex
);
1529 void *map
= upload_state(ice
->state
.surface_uploader
, &isv
->surface_state
,
1530 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
1534 struct iris_bo
*state_bo
= iris_resource_bo(isv
->surface_state
.res
);
1535 isv
->surface_state
.offset
+= iris_bo_offset_from_base_address(state_bo
);
1537 if (util_format_is_depth_or_stencil(tmpl
->format
)) {
1538 struct iris_resource
*zres
, *sres
;
1539 const struct util_format_description
*desc
=
1540 util_format_description(tmpl
->format
);
1542 iris_get_depth_stencil_resources(tex
, &zres
, &sres
);
1544 tex
= util_format_has_depth(desc
) ? &zres
->base
: &sres
->base
;
1547 isv
->res
= (struct iris_resource
*) tex
;
1549 isl_surf_usage_flags_t usage
= ISL_SURF_USAGE_TEXTURE_BIT
;
1551 if (isv
->base
.target
== PIPE_TEXTURE_CUBE
||
1552 isv
->base
.target
== PIPE_TEXTURE_CUBE_ARRAY
)
1553 usage
|= ISL_SURF_USAGE_CUBE_BIT
;
1555 const struct iris_format_info fmt
=
1556 iris_format_for_usage(devinfo
, tmpl
->format
, usage
);
1558 isv
->view
= (struct isl_view
) {
1560 .swizzle
= (struct isl_swizzle
) {
1561 .r
= fmt_swizzle(&fmt
, tmpl
->swizzle_r
),
1562 .g
= fmt_swizzle(&fmt
, tmpl
->swizzle_g
),
1563 .b
= fmt_swizzle(&fmt
, tmpl
->swizzle_b
),
1564 .a
= fmt_swizzle(&fmt
, tmpl
->swizzle_a
),
1569 /* Fill out SURFACE_STATE for this view. */
1570 if (tmpl
->target
!= PIPE_BUFFER
) {
1571 isv
->view
.base_level
= tmpl
->u
.tex
.first_level
;
1572 isv
->view
.levels
= tmpl
->u
.tex
.last_level
- tmpl
->u
.tex
.first_level
+ 1;
1573 // XXX: do I need to port f9fd0cf4790cb2a530e75d1a2206dbb9d8af7cb2?
1574 isv
->view
.base_array_layer
= tmpl
->u
.tex
.first_layer
;
1575 isv
->view
.array_len
=
1576 tmpl
->u
.tex
.last_layer
- tmpl
->u
.tex
.first_layer
+ 1;
1578 isl_surf_fill_state(&screen
->isl_dev
, map
,
1579 .surf
= &isv
->res
->surf
, .view
= &isv
->view
,
1581 .address
= isv
->res
->bo
->gtt_offset
);
1583 // .clear_color = clear_color,
1585 fill_buffer_surface_state(&screen
->isl_dev
, isv
->res
->bo
, map
,
1586 isv
->view
.format
, tmpl
->u
.buf
.offset
,
1594 iris_sampler_view_destroy(struct pipe_context
*ctx
,
1595 struct pipe_sampler_view
*state
)
1597 struct iris_sampler_view
*isv
= (void *) state
;
1598 pipe_resource_reference(&state
->texture
, NULL
);
1599 pipe_resource_reference(&isv
->surface_state
.res
, NULL
);
1604 * The pipe->create_surface() driver hook.
1606 * In Gallium nomenclature, "surfaces" are a view of a resource that
1607 * can be bound as a render target or depth/stencil buffer.
1609 static struct pipe_surface
*
1610 iris_create_surface(struct pipe_context
*ctx
,
1611 struct pipe_resource
*tex
,
1612 const struct pipe_surface
*tmpl
)
1614 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1615 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1616 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1617 struct iris_surface
*surf
= calloc(1, sizeof(struct iris_surface
));
1618 struct pipe_surface
*psurf
= &surf
->base
;
1619 struct iris_resource
*res
= (struct iris_resource
*) tex
;
1624 pipe_reference_init(&psurf
->reference
, 1);
1625 pipe_resource_reference(&psurf
->texture
, tex
);
1626 psurf
->context
= ctx
;
1627 psurf
->format
= tmpl
->format
;
1628 psurf
->width
= tex
->width0
;
1629 psurf
->height
= tex
->height0
;
1630 psurf
->texture
= tex
;
1631 psurf
->u
.tex
.first_layer
= tmpl
->u
.tex
.first_layer
;
1632 psurf
->u
.tex
.last_layer
= tmpl
->u
.tex
.last_layer
;
1633 psurf
->u
.tex
.level
= tmpl
->u
.tex
.level
;
1635 isl_surf_usage_flags_t usage
= 0;
1637 usage
= ISL_SURF_USAGE_STORAGE_BIT
;
1638 else if (util_format_is_depth_or_stencil(tmpl
->format
))
1639 usage
= ISL_SURF_USAGE_DEPTH_BIT
;
1641 usage
= ISL_SURF_USAGE_RENDER_TARGET_BIT
;
1643 const struct iris_format_info fmt
=
1644 iris_format_for_usage(devinfo
, psurf
->format
, usage
);
1646 if ((usage
& ISL_SURF_USAGE_RENDER_TARGET_BIT
) &&
1647 !isl_format_supports_rendering(devinfo
, fmt
.fmt
)) {
1648 /* Framebuffer validation will reject this invalid case, but it
1649 * hasn't had the opportunity yet. In the meantime, we need to
1650 * avoid hitting ISL asserts about unsupported formats below.
1656 surf
->view
= (struct isl_view
) {
1658 .base_level
= tmpl
->u
.tex
.level
,
1660 .base_array_layer
= tmpl
->u
.tex
.first_layer
,
1661 .array_len
= tmpl
->u
.tex
.last_layer
- tmpl
->u
.tex
.first_layer
+ 1,
1662 .swizzle
= ISL_SWIZZLE_IDENTITY
,
1666 /* Bail early for depth/stencil - we don't want SURFACE_STATE for them. */
1667 if (res
->surf
.usage
& (ISL_SURF_USAGE_DEPTH_BIT
|
1668 ISL_SURF_USAGE_STENCIL_BIT
))
1672 void *map
= upload_state(ice
->state
.surface_uploader
, &surf
->surface_state
,
1673 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
1677 struct iris_bo
*state_bo
= iris_resource_bo(surf
->surface_state
.res
);
1678 surf
->surface_state
.offset
+= iris_bo_offset_from_base_address(state_bo
);
1680 isl_surf_fill_state(&screen
->isl_dev
, map
,
1681 .surf
= &res
->surf
, .view
= &surf
->view
,
1683 .address
= res
->bo
->gtt_offset
);
1685 // .clear_color = clear_color,
1691 * The pipe->set_shader_images() driver hook.
1694 iris_set_shader_images(struct pipe_context
*ctx
,
1695 enum pipe_shader_type p_stage
,
1696 unsigned start_slot
, unsigned count
,
1697 const struct pipe_image_view
*p_images
)
1699 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1700 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1701 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1702 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1703 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1705 shs
->bound_image_views
&= ~u_bit_consecutive(start_slot
, count
);
1707 for (unsigned i
= 0; i
< count
; i
++) {
1708 if (p_images
&& p_images
[i
].resource
) {
1709 const struct pipe_image_view
*img
= &p_images
[i
];
1710 struct iris_resource
*res
= (void *) img
->resource
;
1711 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, &res
->base
);
1713 shs
->bound_image_views
|= 1 << (start_slot
+ i
);
1715 res
->bind_history
|= PIPE_BIND_SHADER_IMAGE
;
1717 // XXX: these are not retained forever, use a separate uploader?
1719 upload_state(ice
->state
.surface_uploader
,
1720 &shs
->image
[start_slot
+ i
].surface_state
,
1721 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
1722 if (!unlikely(map
)) {
1723 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, NULL
);
1727 struct iris_bo
*surf_state_bo
=
1728 iris_resource_bo(shs
->image
[start_slot
+ i
].surface_state
.res
);
1729 shs
->image
[start_slot
+ i
].surface_state
.offset
+=
1730 iris_bo_offset_from_base_address(surf_state_bo
);
1732 isl_surf_usage_flags_t usage
= ISL_SURF_USAGE_STORAGE_BIT
;
1733 enum isl_format isl_format
=
1734 iris_format_for_usage(devinfo
, img
->format
, usage
).fmt
;
1736 if (img
->shader_access
& PIPE_IMAGE_ACCESS_READ
)
1737 isl_format
= isl_lower_storage_image_format(devinfo
, isl_format
);
1739 shs
->image
[start_slot
+ i
].access
= img
->shader_access
;
1741 if (res
->base
.target
!= PIPE_BUFFER
) {
1742 struct isl_view view
= {
1743 .format
= isl_format
,
1744 .base_level
= img
->u
.tex
.level
,
1746 .base_array_layer
= img
->u
.tex
.first_layer
,
1747 .array_len
= img
->u
.tex
.last_layer
- img
->u
.tex
.first_layer
+ 1,
1748 .swizzle
= ISL_SWIZZLE_IDENTITY
,
1752 isl_surf_fill_state(&screen
->isl_dev
, map
,
1753 .surf
= &res
->surf
, .view
= &view
,
1755 .address
= res
->bo
->gtt_offset
);
1757 // .clear_color = clear_color,
1759 fill_buffer_surface_state(&screen
->isl_dev
, res
->bo
, map
,
1760 isl_format
, img
->u
.buf
.offset
,
1764 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, NULL
);
1765 pipe_resource_reference(&shs
->image
[start_slot
+ i
].surface_state
.res
,
1770 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
1775 * The pipe->set_sampler_views() driver hook.
1778 iris_set_sampler_views(struct pipe_context
*ctx
,
1779 enum pipe_shader_type p_stage
,
1780 unsigned start
, unsigned count
,
1781 struct pipe_sampler_view
**views
)
1783 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1784 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1785 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1787 shs
->bound_sampler_views
&= ~u_bit_consecutive(start
, count
);
1789 for (unsigned i
= 0; i
< count
; i
++) {
1790 pipe_sampler_view_reference((struct pipe_sampler_view
**)
1791 &shs
->textures
[start
+ i
], views
[i
]);
1792 struct iris_sampler_view
*view
= (void *) views
[i
];
1794 view
->res
->bind_history
|= PIPE_BIND_SAMPLER_VIEW
;
1795 shs
->bound_sampler_views
|= 1 << (start
+ i
);
1799 ice
->state
.dirty
|= (IRIS_DIRTY_BINDINGS_VS
<< stage
);
1803 * The pipe->set_tess_state() driver hook.
1806 iris_set_tess_state(struct pipe_context
*ctx
,
1807 const float default_outer_level
[4],
1808 const float default_inner_level
[2])
1810 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1812 memcpy(&ice
->state
.default_outer_level
[0], &default_outer_level
[0], 4 * sizeof(float));
1813 memcpy(&ice
->state
.default_inner_level
[0], &default_inner_level
[0], 2 * sizeof(float));
1815 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_TCS
;
1819 iris_surface_destroy(struct pipe_context
*ctx
, struct pipe_surface
*p_surf
)
1821 struct iris_surface
*surf
= (void *) p_surf
;
1822 pipe_resource_reference(&p_surf
->texture
, NULL
);
1823 pipe_resource_reference(&surf
->surface_state
.res
, NULL
);
1828 iris_set_clip_state(struct pipe_context
*ctx
,
1829 const struct pipe_clip_state
*state
)
1831 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1832 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_VERTEX
];
1834 memcpy(&ice
->state
.clip_planes
, state
, sizeof(*state
));
1836 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
;
1837 shs
->cbuf0_needs_upload
= true;
1841 * The pipe->set_polygon_stipple() driver hook.
1844 iris_set_polygon_stipple(struct pipe_context
*ctx
,
1845 const struct pipe_poly_stipple
*state
)
1847 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1848 memcpy(&ice
->state
.poly_stipple
, state
, sizeof(*state
));
1849 ice
->state
.dirty
|= IRIS_DIRTY_POLYGON_STIPPLE
;
1853 * The pipe->set_sample_mask() driver hook.
1856 iris_set_sample_mask(struct pipe_context
*ctx
, unsigned sample_mask
)
1858 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1860 /* We only support 16x MSAA, so we have 16 bits of sample maks.
1861 * st/mesa may pass us 0xffffffff though, meaning "enable all samples".
1863 ice
->state
.sample_mask
= sample_mask
& 0xffff;
1864 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLE_MASK
;
1868 * The pipe->set_scissor_states() driver hook.
1870 * This corresponds to our SCISSOR_RECT state structures. It's an
1871 * exact match, so we just store them, and memcpy them out later.
1874 iris_set_scissor_states(struct pipe_context
*ctx
,
1875 unsigned start_slot
,
1876 unsigned num_scissors
,
1877 const struct pipe_scissor_state
*rects
)
1879 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1881 for (unsigned i
= 0; i
< num_scissors
; i
++) {
1882 if (rects
[i
].minx
== rects
[i
].maxx
|| rects
[i
].miny
== rects
[i
].maxy
) {
1883 /* If the scissor was out of bounds and got clamped to 0 width/height
1884 * at the bounds, the subtraction of 1 from maximums could produce a
1885 * negative number and thus not clip anything. Instead, just provide
1886 * a min > max scissor inside the bounds, which produces the expected
1889 ice
->state
.scissors
[start_slot
+ i
] = (struct pipe_scissor_state
) {
1890 .minx
= 1, .maxx
= 0, .miny
= 1, .maxy
= 0,
1893 ice
->state
.scissors
[start_slot
+ i
] = (struct pipe_scissor_state
) {
1894 .minx
= rects
[i
].minx
, .miny
= rects
[i
].miny
,
1895 .maxx
= rects
[i
].maxx
- 1, .maxy
= rects
[i
].maxy
- 1,
1900 ice
->state
.dirty
|= IRIS_DIRTY_SCISSOR_RECT
;
1904 * The pipe->set_stencil_ref() driver hook.
1906 * This is added to 3DSTATE_WM_DEPTH_STENCIL dynamically at draw time.
1909 iris_set_stencil_ref(struct pipe_context
*ctx
,
1910 const struct pipe_stencil_ref
*state
)
1912 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1913 memcpy(&ice
->state
.stencil_ref
, state
, sizeof(*state
));
1914 ice
->state
.dirty
|= IRIS_DIRTY_WM_DEPTH_STENCIL
;
1918 viewport_extent(const struct pipe_viewport_state
*state
, int axis
, float sign
)
1920 return copysignf(state
->scale
[axis
], sign
) + state
->translate
[axis
];
1924 calculate_guardband_size(uint32_t fb_width
, uint32_t fb_height
,
1925 float m00
, float m11
, float m30
, float m31
,
1926 float *xmin
, float *xmax
,
1927 float *ymin
, float *ymax
)
1929 /* According to the "Vertex X,Y Clamping and Quantization" section of the
1930 * Strips and Fans documentation:
1932 * "The vertex X and Y screen-space coordinates are also /clamped/ to the
1933 * fixed-point "guardband" range supported by the rasterization hardware"
1937 * "In almost all circumstances, if an object’s vertices are actually
1938 * modified by this clamping (i.e., had X or Y coordinates outside of
1939 * the guardband extent the rendered object will not match the intended
1940 * result. Therefore software should take steps to ensure that this does
1941 * not happen - e.g., by clipping objects such that they do not exceed
1942 * these limits after the Drawing Rectangle is applied."
1944 * I believe the fundamental restriction is that the rasterizer (in
1945 * the SF/WM stages) have a limit on the number of pixels that can be
1946 * rasterized. We need to ensure any coordinates beyond the rasterizer
1947 * limit are handled by the clipper. So effectively that limit becomes
1948 * the clipper's guardband size.
1950 * It goes on to say:
1952 * "In addition, in order to be correctly rendered, objects must have a
1953 * screenspace bounding box not exceeding 8K in the X or Y direction.
1954 * This additional restriction must also be comprehended by software,
1955 * i.e., enforced by use of clipping."
1957 * This makes no sense. Gen7+ hardware supports 16K render targets,
1958 * and you definitely need to be able to draw polygons that fill the
1959 * surface. Our assumption is that the rasterizer was limited to 8K
1960 * on Sandybridge, which only supports 8K surfaces, and it was actually
1961 * increased to 16K on Ivybridge and later.
1963 * So, limit the guardband to 16K on Gen7+ and 8K on Sandybridge.
1965 const float gb_size
= GEN_GEN
>= 7 ? 16384.0f
: 8192.0f
;
1967 if (m00
!= 0 && m11
!= 0) {
1968 /* First, we compute the screen-space render area */
1969 const float ss_ra_xmin
= MIN3( 0, m30
+ m00
, m30
- m00
);
1970 const float ss_ra_xmax
= MAX3( fb_width
, m30
+ m00
, m30
- m00
);
1971 const float ss_ra_ymin
= MIN3( 0, m31
+ m11
, m31
- m11
);
1972 const float ss_ra_ymax
= MAX3(fb_height
, m31
+ m11
, m31
- m11
);
1974 /* We want the guardband to be centered on that */
1975 const float ss_gb_xmin
= (ss_ra_xmin
+ ss_ra_xmax
) / 2 - gb_size
;
1976 const float ss_gb_xmax
= (ss_ra_xmin
+ ss_ra_xmax
) / 2 + gb_size
;
1977 const float ss_gb_ymin
= (ss_ra_ymin
+ ss_ra_ymax
) / 2 - gb_size
;
1978 const float ss_gb_ymax
= (ss_ra_ymin
+ ss_ra_ymax
) / 2 + gb_size
;
1980 /* Now we need it in native device coordinates */
1981 const float ndc_gb_xmin
= (ss_gb_xmin
- m30
) / m00
;
1982 const float ndc_gb_xmax
= (ss_gb_xmax
- m30
) / m00
;
1983 const float ndc_gb_ymin
= (ss_gb_ymin
- m31
) / m11
;
1984 const float ndc_gb_ymax
= (ss_gb_ymax
- m31
) / m11
;
1986 /* Thanks to Y-flipping and ORIGIN_UPPER_LEFT, the Y coordinates may be
1987 * flipped upside-down. X should be fine though.
1989 assert(ndc_gb_xmin
<= ndc_gb_xmax
);
1990 *xmin
= ndc_gb_xmin
;
1991 *xmax
= ndc_gb_xmax
;
1992 *ymin
= MIN2(ndc_gb_ymin
, ndc_gb_ymax
);
1993 *ymax
= MAX2(ndc_gb_ymin
, ndc_gb_ymax
);
1995 /* The viewport scales to 0, so nothing will be rendered. */
2004 * The pipe->set_viewport_states() driver hook.
2006 * This corresponds to our SF_CLIP_VIEWPORT states. We can't calculate
2007 * the guardband yet, as we need the framebuffer dimensions, but we can
2008 * at least fill out the rest.
2011 iris_set_viewport_states(struct pipe_context
*ctx
,
2012 unsigned start_slot
,
2014 const struct pipe_viewport_state
*states
)
2016 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2018 memcpy(&ice
->state
.viewports
[start_slot
], states
, sizeof(*states
) * count
);
2020 ice
->state
.dirty
|= IRIS_DIRTY_SF_CL_VIEWPORT
;
2022 if (ice
->state
.cso_rast
&& (!ice
->state
.cso_rast
->depth_clip_near
||
2023 !ice
->state
.cso_rast
->depth_clip_far
))
2024 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
2028 * The pipe->set_framebuffer_state() driver hook.
2030 * Sets the current draw FBO, including color render targets, depth,
2031 * and stencil buffers.
2034 iris_set_framebuffer_state(struct pipe_context
*ctx
,
2035 const struct pipe_framebuffer_state
*state
)
2037 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2038 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2039 struct isl_device
*isl_dev
= &screen
->isl_dev
;
2040 struct pipe_framebuffer_state
*cso
= &ice
->state
.framebuffer
;
2041 struct iris_resource
*zres
;
2042 struct iris_resource
*stencil_res
;
2044 unsigned samples
= util_framebuffer_get_num_samples(state
);
2046 if (cso
->samples
!= samples
) {
2047 ice
->state
.dirty
|= IRIS_DIRTY_MULTISAMPLE
;
2050 if (cso
->nr_cbufs
!= state
->nr_cbufs
) {
2051 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
2054 if ((cso
->layers
== 0) != (state
->layers
== 0)) {
2055 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
2058 if (cso
->width
!= state
->width
|| cso
->height
!= state
->height
) {
2059 ice
->state
.dirty
|= IRIS_DIRTY_SF_CL_VIEWPORT
;
2062 util_copy_framebuffer_state(cso
, state
);
2063 cso
->samples
= samples
;
2065 struct iris_depth_buffer_state
*cso_z
= &ice
->state
.genx
->depth_buffer
;
2067 struct isl_view view
= {
2070 .base_array_layer
= 0,
2072 .swizzle
= ISL_SWIZZLE_IDENTITY
,
2075 struct isl_depth_stencil_hiz_emit_info info
= {
2081 iris_get_depth_stencil_resources(cso
->zsbuf
->texture
, &zres
,
2084 view
.base_level
= cso
->zsbuf
->u
.tex
.level
;
2085 view
.base_array_layer
= cso
->zsbuf
->u
.tex
.first_layer
;
2087 cso
->zsbuf
->u
.tex
.last_layer
- cso
->zsbuf
->u
.tex
.first_layer
+ 1;
2090 view
.usage
|= ISL_SURF_USAGE_DEPTH_BIT
;
2092 info
.depth_surf
= &zres
->surf
;
2093 info
.depth_address
= zres
->bo
->gtt_offset
;
2094 info
.hiz_usage
= ISL_AUX_USAGE_NONE
;
2096 view
.format
= zres
->surf
.format
;
2100 view
.usage
|= ISL_SURF_USAGE_STENCIL_BIT
;
2101 info
.stencil_surf
= &stencil_res
->surf
;
2102 info
.stencil_address
= stencil_res
->bo
->gtt_offset
;
2104 view
.format
= stencil_res
->surf
.format
;
2108 isl_emit_depth_stencil_hiz_s(isl_dev
, cso_z
->packets
, &info
);
2110 /* Make a null surface for unbound buffers */
2111 void *null_surf_map
=
2112 upload_state(ice
->state
.surface_uploader
, &ice
->state
.null_fb
,
2113 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2114 isl_null_fill_state(&screen
->isl_dev
, null_surf_map
,
2115 isl_extent3d(MAX2(cso
->width
, 1),
2116 MAX2(cso
->height
, 1),
2117 cso
->layers
? cso
->layers
: 1));
2118 ice
->state
.null_fb
.offset
+=
2119 iris_bo_offset_from_base_address(iris_resource_bo(ice
->state
.null_fb
.res
));
2121 ice
->state
.dirty
|= IRIS_DIRTY_DEPTH_BUFFER
;
2123 /* Render target change */
2124 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_FS
;
2126 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_FRAMEBUFFER
];
2129 // XXX: we may want to flag IRIS_DIRTY_MULTISAMPLE (or SAMPLE_MASK?)
2130 // XXX: see commit 979fc1bc9bcc64027ff2cfafd285676f31b930a6
2132 /* The PIPE_CONTROL command description says:
2134 * "Whenever a Binding Table Index (BTI) used by a Render Target Message
2135 * points to a different RENDER_SURFACE_STATE, SW must issue a Render
2136 * Target Cache Flush by enabling this bit. When render target flush
2137 * is set due to new association of BTI, PS Scoreboard Stall bit must
2138 * be set in this packet."
2140 // XXX: does this need to happen at 3DSTATE_BTP_PS time?
2141 iris_emit_pipe_control_flush(&ice
->batches
[IRIS_BATCH_RENDER
],
2142 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
2143 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
2148 upload_ubo_surf_state(struct iris_context
*ice
,
2149 struct iris_const_buffer
*cbuf
,
2150 unsigned buffer_size
)
2152 struct pipe_context
*ctx
= &ice
->ctx
;
2153 struct iris_screen
*screen
= (struct iris_screen
*) ctx
->screen
;
2155 // XXX: these are not retained forever, use a separate uploader?
2157 upload_state(ice
->state
.surface_uploader
, &cbuf
->surface_state
,
2158 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2159 if (!unlikely(map
)) {
2160 pipe_resource_reference(&cbuf
->data
.res
, NULL
);
2164 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
2165 struct iris_bo
*surf_bo
= iris_resource_bo(cbuf
->surface_state
.res
);
2166 cbuf
->surface_state
.offset
+= iris_bo_offset_from_base_address(surf_bo
);
2168 isl_buffer_fill_state(&screen
->isl_dev
, map
,
2169 .address
= res
->bo
->gtt_offset
+ cbuf
->data
.offset
,
2170 .size_B
= MIN2(buffer_size
,
2171 res
->bo
->size
- cbuf
->data
.offset
),
2172 .format
= ISL_FORMAT_R32G32B32A32_FLOAT
,
2178 * The pipe->set_constant_buffer() driver hook.
2180 * This uploads any constant data in user buffers, and references
2181 * any UBO resources containing constant data.
2184 iris_set_constant_buffer(struct pipe_context
*ctx
,
2185 enum pipe_shader_type p_stage
, unsigned index
,
2186 const struct pipe_constant_buffer
*input
)
2188 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2189 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2190 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2191 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[index
];
2193 if (input
&& input
->buffer
) {
2196 pipe_resource_reference(&cbuf
->data
.res
, input
->buffer
);
2197 cbuf
->data
.offset
= input
->buffer_offset
;
2199 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
2200 res
->bind_history
|= PIPE_BIND_CONSTANT_BUFFER
;
2202 upload_ubo_surf_state(ice
, cbuf
, input
->buffer_size
);
2204 pipe_resource_reference(&cbuf
->data
.res
, NULL
);
2205 pipe_resource_reference(&cbuf
->surface_state
.res
, NULL
);
2210 memcpy(&shs
->cbuf0
, input
, sizeof(shs
->cbuf0
));
2212 memset(&shs
->cbuf0
, 0, sizeof(shs
->cbuf0
));
2214 shs
->cbuf0_needs_upload
= true;
2217 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
<< stage
;
2218 // XXX: maybe not necessary all the time...?
2219 // XXX: we need 3DS_BTP to commit these changes, and if we fell back to
2220 // XXX: pull model we may need actual new bindings...
2221 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2225 upload_uniforms(struct iris_context
*ice
,
2226 gl_shader_stage stage
)
2228 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2229 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[0];
2230 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
2232 unsigned upload_size
= shader
->num_system_values
* sizeof(uint32_t) +
2233 shs
->cbuf0
.buffer_size
;
2235 if (upload_size
== 0)
2239 upload_state(ice
->ctx
.const_uploader
, &cbuf
->data
, upload_size
, 64);
2241 for (int i
= 0; i
< shader
->num_system_values
; i
++) {
2242 uint32_t sysval
= shader
->system_values
[i
];
2245 if (BRW_PARAM_BUILTIN_IS_CLIP_PLANE(sysval
)) {
2246 int plane
= BRW_PARAM_BUILTIN_CLIP_PLANE_IDX(sysval
);
2247 int comp
= BRW_PARAM_BUILTIN_CLIP_PLANE_COMP(sysval
);
2248 value
= fui(ice
->state
.clip_planes
.ucp
[plane
][comp
]);
2249 } else if (sysval
== BRW_PARAM_BUILTIN_PATCH_VERTICES_IN
) {
2250 if (stage
== MESA_SHADER_TESS_CTRL
) {
2251 value
= ice
->state
.vertices_per_patch
;
2253 assert(stage
== MESA_SHADER_TESS_EVAL
);
2254 const struct shader_info
*tcs_info
=
2255 iris_get_shader_info(ice
, MESA_SHADER_TESS_CTRL
);
2258 value
= tcs_info
->tess
.tcs_vertices_out
;
2261 assert(!"unhandled system value");
2267 if (shs
->cbuf0
.user_buffer
) {
2268 memcpy(map
, shs
->cbuf0
.user_buffer
, shs
->cbuf0
.buffer_size
);
2271 upload_ubo_surf_state(ice
, cbuf
, upload_size
);
2275 * The pipe->set_shader_buffers() driver hook.
2277 * This binds SSBOs and ABOs. Unfortunately, we need to stream out
2278 * SURFACE_STATE here, as the buffer offset may change each time.
2281 iris_set_shader_buffers(struct pipe_context
*ctx
,
2282 enum pipe_shader_type p_stage
,
2283 unsigned start_slot
, unsigned count
,
2284 const struct pipe_shader_buffer
*buffers
)
2286 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2287 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2288 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2289 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2291 for (unsigned i
= 0; i
< count
; i
++) {
2292 if (buffers
&& buffers
[i
].buffer
) {
2293 const struct pipe_shader_buffer
*buffer
= &buffers
[i
];
2294 struct iris_resource
*res
= (void *) buffer
->buffer
;
2295 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], &res
->base
);
2297 res
->bind_history
|= PIPE_BIND_SHADER_BUFFER
;
2299 // XXX: these are not retained forever, use a separate uploader?
2301 upload_state(ice
->state
.surface_uploader
,
2302 &shs
->ssbo_surface_state
[start_slot
+ i
],
2303 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2304 if (!unlikely(map
)) {
2305 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], NULL
);
2309 struct iris_bo
*surf_state_bo
=
2310 iris_resource_bo(shs
->ssbo_surface_state
[start_slot
+ i
].res
);
2311 shs
->ssbo_surface_state
[start_slot
+ i
].offset
+=
2312 iris_bo_offset_from_base_address(surf_state_bo
);
2314 isl_buffer_fill_state(&screen
->isl_dev
, map
,
2316 res
->bo
->gtt_offset
+ buffer
->buffer_offset
,
2318 MIN2(buffer
->buffer_size
,
2319 res
->bo
->size
- buffer
->buffer_offset
),
2320 .format
= ISL_FORMAT_RAW
,
2324 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], NULL
);
2325 pipe_resource_reference(&shs
->ssbo_surface_state
[start_slot
+ i
].res
,
2330 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2334 iris_delete_state(struct pipe_context
*ctx
, void *state
)
2340 * The pipe->set_vertex_buffers() driver hook.
2342 * This translates pipe_vertex_buffer to our 3DSTATE_VERTEX_BUFFERS packet.
2345 iris_set_vertex_buffers(struct pipe_context
*ctx
,
2346 unsigned start_slot
, unsigned count
,
2347 const struct pipe_vertex_buffer
*buffers
)
2349 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2350 struct iris_genx_state
*genx
= ice
->state
.genx
;
2352 ice
->state
.bound_vertex_buffers
&= ~u_bit_consecutive64(start_slot
, count
);
2354 for (unsigned i
= 0; i
< count
; i
++) {
2355 const struct pipe_vertex_buffer
*buffer
= buffers
? &buffers
[i
] : NULL
;
2356 struct iris_vertex_buffer_state
*state
=
2357 &genx
->vertex_buffers
[start_slot
+ i
];
2360 pipe_resource_reference(&state
->resource
, NULL
);
2364 assert(!buffer
->is_user_buffer
);
2366 ice
->state
.bound_vertex_buffers
|= 1ull << (start_slot
+ i
);
2368 pipe_resource_reference(&state
->resource
, buffer
->buffer
.resource
);
2369 struct iris_resource
*res
= (void *) state
->resource
;
2372 res
->bind_history
|= PIPE_BIND_VERTEX_BUFFER
;
2374 iris_pack_state(GENX(VERTEX_BUFFER_STATE
), state
->state
, vb
) {
2375 vb
.VertexBufferIndex
= start_slot
+ i
;
2377 vb
.AddressModifyEnable
= true;
2378 vb
.BufferPitch
= buffer
->stride
;
2380 vb
.BufferSize
= res
->bo
->size
;
2381 vb
.BufferStartingAddress
=
2382 ro_bo(NULL
, res
->bo
->gtt_offset
+ buffer
->buffer_offset
);
2384 vb
.NullVertexBuffer
= true;
2389 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_BUFFERS
;
2393 * Gallium CSO for vertex elements.
2395 struct iris_vertex_element_state
{
2396 uint32_t vertex_elements
[1 + 33 * GENX(VERTEX_ELEMENT_STATE_length
)];
2397 uint32_t vf_instancing
[33 * GENX(3DSTATE_VF_INSTANCING_length
)];
2402 * The pipe->create_vertex_elements() driver hook.
2404 * This translates pipe_vertex_element to our 3DSTATE_VERTEX_ELEMENTS
2405 * and 3DSTATE_VF_INSTANCING commands. SGVs are handled at draw time.
2408 iris_create_vertex_elements(struct pipe_context
*ctx
,
2410 const struct pipe_vertex_element
*state
)
2412 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2413 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2414 struct iris_vertex_element_state
*cso
=
2415 malloc(sizeof(struct iris_vertex_element_state
));
2420 * - create edge flag one
2422 * - if those are necessary, use count + 1/2/3... OR in the length
2424 iris_pack_command(GENX(3DSTATE_VERTEX_ELEMENTS
), cso
->vertex_elements
, ve
) {
2426 1 + GENX(VERTEX_ELEMENT_STATE_length
) * MAX2(count
, 1) - 2;
2429 uint32_t *ve_pack_dest
= &cso
->vertex_elements
[1];
2430 uint32_t *vfi_pack_dest
= cso
->vf_instancing
;
2433 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
2435 ve
.SourceElementFormat
= ISL_FORMAT_R32G32B32A32_FLOAT
;
2436 ve
.Component0Control
= VFCOMP_STORE_0
;
2437 ve
.Component1Control
= VFCOMP_STORE_0
;
2438 ve
.Component2Control
= VFCOMP_STORE_0
;
2439 ve
.Component3Control
= VFCOMP_STORE_1_FP
;
2442 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
2446 for (int i
= 0; i
< count
; i
++) {
2447 const struct iris_format_info fmt
=
2448 iris_format_for_usage(devinfo
, state
[i
].src_format
, 0);
2449 unsigned comp
[4] = { VFCOMP_STORE_SRC
, VFCOMP_STORE_SRC
,
2450 VFCOMP_STORE_SRC
, VFCOMP_STORE_SRC
};
2452 switch (isl_format_get_num_channels(fmt
.fmt
)) {
2453 case 0: comp
[0] = VFCOMP_STORE_0
;
2454 case 1: comp
[1] = VFCOMP_STORE_0
;
2455 case 2: comp
[2] = VFCOMP_STORE_0
;
2457 comp
[3] = isl_format_has_int_channel(fmt
.fmt
) ? VFCOMP_STORE_1_INT
2458 : VFCOMP_STORE_1_FP
;
2461 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
2462 ve
.VertexBufferIndex
= state
[i
].vertex_buffer_index
;
2464 ve
.SourceElementOffset
= state
[i
].src_offset
;
2465 ve
.SourceElementFormat
= fmt
.fmt
;
2466 ve
.Component0Control
= comp
[0];
2467 ve
.Component1Control
= comp
[1];
2468 ve
.Component2Control
= comp
[2];
2469 ve
.Component3Control
= comp
[3];
2472 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
2473 vi
.VertexElementIndex
= i
;
2474 vi
.InstancingEnable
= state
[i
].instance_divisor
> 0;
2475 vi
.InstanceDataStepRate
= state
[i
].instance_divisor
;
2478 ve_pack_dest
+= GENX(VERTEX_ELEMENT_STATE_length
);
2479 vfi_pack_dest
+= GENX(3DSTATE_VF_INSTANCING_length
);
2486 * The pipe->bind_vertex_elements_state() driver hook.
2489 iris_bind_vertex_elements_state(struct pipe_context
*ctx
, void *state
)
2491 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2492 struct iris_vertex_element_state
*old_cso
= ice
->state
.cso_vertex_elements
;
2493 struct iris_vertex_element_state
*new_cso
= state
;
2495 /* 3DSTATE_VF_SGVs overrides the last VE, so if the count is changing,
2496 * we need to re-emit it to ensure we're overriding the right one.
2498 if (new_cso
&& cso_changed(count
))
2499 ice
->state
.dirty
|= IRIS_DIRTY_VF_SGVS
;
2501 ice
->state
.cso_vertex_elements
= state
;
2502 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_ELEMENTS
;
2506 * The pipe->create_stream_output_target() driver hook.
2508 * "Target" here refers to a destination buffer. We translate this into
2509 * a 3DSTATE_SO_BUFFER packet. We can handle most fields, but don't yet
2510 * know which buffer this represents, or whether we ought to zero the
2511 * write-offsets, or append. Those are handled in the set() hook.
2513 static struct pipe_stream_output_target
*
2514 iris_create_stream_output_target(struct pipe_context
*ctx
,
2515 struct pipe_resource
*p_res
,
2516 unsigned buffer_offset
,
2517 unsigned buffer_size
)
2519 struct iris_resource
*res
= (void *) p_res
;
2520 struct iris_stream_output_target
*cso
= calloc(1, sizeof(*cso
));
2524 res
->bind_history
|= PIPE_BIND_STREAM_OUTPUT
;
2526 pipe_reference_init(&cso
->base
.reference
, 1);
2527 pipe_resource_reference(&cso
->base
.buffer
, p_res
);
2528 cso
->base
.buffer_offset
= buffer_offset
;
2529 cso
->base
.buffer_size
= buffer_size
;
2530 cso
->base
.context
= ctx
;
2532 upload_state(ctx
->stream_uploader
, &cso
->offset
, sizeof(uint32_t), 4);
2538 iris_stream_output_target_destroy(struct pipe_context
*ctx
,
2539 struct pipe_stream_output_target
*state
)
2541 struct iris_stream_output_target
*cso
= (void *) state
;
2543 pipe_resource_reference(&cso
->base
.buffer
, NULL
);
2544 pipe_resource_reference(&cso
->offset
.res
, NULL
);
2550 * The pipe->set_stream_output_targets() driver hook.
2552 * At this point, we know which targets are bound to a particular index,
2553 * and also whether we want to append or start over. We can finish the
2554 * 3DSTATE_SO_BUFFER packets we started earlier.
2557 iris_set_stream_output_targets(struct pipe_context
*ctx
,
2558 unsigned num_targets
,
2559 struct pipe_stream_output_target
**targets
,
2560 const unsigned *offsets
)
2562 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2563 struct iris_genx_state
*genx
= ice
->state
.genx
;
2564 uint32_t *so_buffers
= genx
->so_buffers
;
2566 const bool active
= num_targets
> 0;
2567 if (ice
->state
.streamout_active
!= active
) {
2568 ice
->state
.streamout_active
= active
;
2569 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
;
2571 /* We only emit 3DSTATE_SO_DECL_LIST when streamout is active, because
2572 * it's a non-pipelined command. If we're switching streamout on, we
2573 * may have missed emitting it earlier, so do so now. (We're already
2574 * taking a stall to update 3DSTATE_SO_BUFFERS anyway...)
2577 ice
->state
.dirty
|= IRIS_DIRTY_SO_DECL_LIST
;
2580 for (int i
= 0; i
< 4; i
++) {
2581 pipe_so_target_reference(&ice
->state
.so_target
[i
],
2582 i
< num_targets
? targets
[i
] : NULL
);
2585 /* No need to update 3DSTATE_SO_BUFFER unless SOL is active. */
2589 for (unsigned i
= 0; i
< 4; i
++,
2590 so_buffers
+= GENX(3DSTATE_SO_BUFFER_length
)) {
2592 if (i
>= num_targets
|| !targets
[i
]) {
2593 iris_pack_command(GENX(3DSTATE_SO_BUFFER
), so_buffers
, sob
)
2594 sob
.SOBufferIndex
= i
;
2598 struct iris_stream_output_target
*tgt
= (void *) targets
[i
];
2599 struct iris_resource
*res
= (void *) tgt
->base
.buffer
;
2601 /* Note that offsets[i] will either be 0, causing us to zero
2602 * the value in the buffer, or 0xFFFFFFFF, which happens to mean
2603 * "continue appending at the existing offset."
2605 assert(offsets
[i
] == 0 || offsets
[i
] == 0xFFFFFFFF);
2607 iris_pack_command(GENX(3DSTATE_SO_BUFFER
), so_buffers
, sob
) {
2608 sob
.SurfaceBaseAddress
=
2609 rw_bo(NULL
, res
->bo
->gtt_offset
+ tgt
->base
.buffer_offset
);
2610 sob
.SOBufferEnable
= true;
2611 sob
.StreamOffsetWriteEnable
= true;
2612 sob
.StreamOutputBufferOffsetAddressEnable
= true;
2613 sob
.MOCS
= MOCS_WB
; // XXX: MOCS
2615 sob
.SurfaceSize
= MAX2(tgt
->base
.buffer_size
/ 4, 1) - 1;
2617 sob
.SOBufferIndex
= i
;
2618 sob
.StreamOffset
= offsets
[i
];
2619 sob
.StreamOutputBufferOffsetAddress
=
2620 rw_bo(NULL
, iris_resource_bo(tgt
->offset
.res
)->gtt_offset
+
2621 tgt
->offset
.offset
);
2625 ice
->state
.dirty
|= IRIS_DIRTY_SO_BUFFERS
;
2629 * An iris-vtable helper for encoding the 3DSTATE_SO_DECL_LIST and
2630 * 3DSTATE_STREAMOUT packets.
2632 * 3DSTATE_SO_DECL_LIST is a list of shader outputs we want the streamout
2633 * hardware to record. We can create it entirely based on the shader, with
2634 * no dynamic state dependencies.
2636 * 3DSTATE_STREAMOUT is an annoying mix of shader-based information and
2637 * state-based settings. We capture the shader-related ones here, and merge
2638 * the rest in at draw time.
2641 iris_create_so_decl_list(const struct pipe_stream_output_info
*info
,
2642 const struct brw_vue_map
*vue_map
)
2644 struct GENX(SO_DECL
) so_decl
[MAX_VERTEX_STREAMS
][128];
2645 int buffer_mask
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2646 int next_offset
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2647 int decls
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2649 STATIC_ASSERT(ARRAY_SIZE(so_decl
[0]) >= MAX_PROGRAM_OUTPUTS
);
2651 memset(so_decl
, 0, sizeof(so_decl
));
2653 /* Construct the list of SO_DECLs to be emitted. The formatting of the
2654 * command feels strange -- each dword pair contains a SO_DECL per stream.
2656 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
2657 const struct pipe_stream_output
*output
= &info
->output
[i
];
2658 const int buffer
= output
->output_buffer
;
2659 const int varying
= output
->register_index
;
2660 const unsigned stream_id
= output
->stream
;
2661 assert(stream_id
< MAX_VERTEX_STREAMS
);
2663 buffer_mask
[stream_id
] |= 1 << buffer
;
2665 assert(vue_map
->varying_to_slot
[varying
] >= 0);
2667 /* Mesa doesn't store entries for gl_SkipComponents in the Outputs[]
2668 * array. Instead, it simply increments DstOffset for the following
2669 * input by the number of components that should be skipped.
2671 * Our hardware is unusual in that it requires us to program SO_DECLs
2672 * for fake "hole" components, rather than simply taking the offset
2673 * for each real varying. Each hole can have size 1, 2, 3, or 4; we
2674 * program as many size = 4 holes as we can, then a final hole to
2675 * accommodate the final 1, 2, or 3 remaining.
2677 int skip_components
= output
->dst_offset
- next_offset
[buffer
];
2679 while (skip_components
> 0) {
2680 so_decl
[stream_id
][decls
[stream_id
]++] = (struct GENX(SO_DECL
)) {
2682 .OutputBufferSlot
= output
->output_buffer
,
2683 .ComponentMask
= (1 << MIN2(skip_components
, 4)) - 1,
2685 skip_components
-= 4;
2688 next_offset
[buffer
] = output
->dst_offset
+ output
->num_components
;
2690 so_decl
[stream_id
][decls
[stream_id
]++] = (struct GENX(SO_DECL
)) {
2691 .OutputBufferSlot
= output
->output_buffer
,
2692 .RegisterIndex
= vue_map
->varying_to_slot
[varying
],
2694 ((1 << output
->num_components
) - 1) << output
->start_component
,
2697 if (decls
[stream_id
] > max_decls
)
2698 max_decls
= decls
[stream_id
];
2701 unsigned dwords
= GENX(3DSTATE_STREAMOUT_length
) + (3 + 2 * max_decls
);
2702 uint32_t *map
= ralloc_size(NULL
, sizeof(uint32_t) * dwords
);
2703 uint32_t *so_decl_map
= map
+ GENX(3DSTATE_STREAMOUT_length
);
2705 iris_pack_command(GENX(3DSTATE_STREAMOUT
), map
, sol
) {
2706 int urb_entry_read_offset
= 0;
2707 int urb_entry_read_length
= (vue_map
->num_slots
+ 1) / 2 -
2708 urb_entry_read_offset
;
2710 /* We always read the whole vertex. This could be reduced at some
2711 * point by reading less and offsetting the register index in the
2714 sol
.Stream0VertexReadOffset
= urb_entry_read_offset
;
2715 sol
.Stream0VertexReadLength
= urb_entry_read_length
- 1;
2716 sol
.Stream1VertexReadOffset
= urb_entry_read_offset
;
2717 sol
.Stream1VertexReadLength
= urb_entry_read_length
- 1;
2718 sol
.Stream2VertexReadOffset
= urb_entry_read_offset
;
2719 sol
.Stream2VertexReadLength
= urb_entry_read_length
- 1;
2720 sol
.Stream3VertexReadOffset
= urb_entry_read_offset
;
2721 sol
.Stream3VertexReadLength
= urb_entry_read_length
- 1;
2723 /* Set buffer pitches; 0 means unbound. */
2724 sol
.Buffer0SurfacePitch
= 4 * info
->stride
[0];
2725 sol
.Buffer1SurfacePitch
= 4 * info
->stride
[1];
2726 sol
.Buffer2SurfacePitch
= 4 * info
->stride
[2];
2727 sol
.Buffer3SurfacePitch
= 4 * info
->stride
[3];
2730 iris_pack_command(GENX(3DSTATE_SO_DECL_LIST
), so_decl_map
, list
) {
2731 list
.DWordLength
= 3 + 2 * max_decls
- 2;
2732 list
.StreamtoBufferSelects0
= buffer_mask
[0];
2733 list
.StreamtoBufferSelects1
= buffer_mask
[1];
2734 list
.StreamtoBufferSelects2
= buffer_mask
[2];
2735 list
.StreamtoBufferSelects3
= buffer_mask
[3];
2736 list
.NumEntries0
= decls
[0];
2737 list
.NumEntries1
= decls
[1];
2738 list
.NumEntries2
= decls
[2];
2739 list
.NumEntries3
= decls
[3];
2742 for (int i
= 0; i
< max_decls
; i
++) {
2743 iris_pack_state(GENX(SO_DECL_ENTRY
), so_decl_map
+ 3 + i
* 2, entry
) {
2744 entry
.Stream0Decl
= so_decl
[0][i
];
2745 entry
.Stream1Decl
= so_decl
[1][i
];
2746 entry
.Stream2Decl
= so_decl
[2][i
];
2747 entry
.Stream3Decl
= so_decl
[3][i
];
2755 iris_compute_sbe_urb_read_interval(uint64_t fs_input_slots
,
2756 const struct brw_vue_map
*last_vue_map
,
2757 bool two_sided_color
,
2758 unsigned *out_offset
,
2759 unsigned *out_length
)
2761 /* The compiler computes the first URB slot without considering COL/BFC
2762 * swizzling (because it doesn't know whether it's enabled), so we need
2763 * to do that here too. This may result in a smaller offset, which
2766 const unsigned first_slot
=
2767 brw_compute_first_urb_slot_required(fs_input_slots
, last_vue_map
);
2769 /* This becomes the URB read offset (counted in pairs of slots). */
2770 assert(first_slot
% 2 == 0);
2771 *out_offset
= first_slot
/ 2;
2773 /* We need to adjust the inputs read to account for front/back color
2774 * swizzling, as it can make the URB length longer.
2776 for (int c
= 0; c
<= 1; c
++) {
2777 if (fs_input_slots
& (VARYING_BIT_COL0
<< c
)) {
2778 /* If two sided color is enabled, the fragment shader's gl_Color
2779 * (COL0) input comes from either the gl_FrontColor (COL0) or
2780 * gl_BackColor (BFC0) input varyings. Mark BFC as used, too.
2782 if (two_sided_color
)
2783 fs_input_slots
|= (VARYING_BIT_BFC0
<< c
);
2785 /* If front color isn't written, we opt to give them back color
2786 * instead of an undefined value. Switch from COL to BFC.
2788 if (last_vue_map
->varying_to_slot
[VARYING_SLOT_COL0
+ c
] == -1) {
2789 fs_input_slots
&= ~(VARYING_BIT_COL0
<< c
);
2790 fs_input_slots
|= (VARYING_BIT_BFC0
<< c
);
2795 /* Compute the minimum URB Read Length necessary for the FS inputs.
2797 * From the Sandy Bridge PRM, Volume 2, Part 1, documentation for
2798 * 3DSTATE_SF DWord 1 bits 15:11, "Vertex URB Entry Read Length":
2800 * "This field should be set to the minimum length required to read the
2801 * maximum source attribute. The maximum source attribute is indicated
2802 * by the maximum value of the enabled Attribute # Source Attribute if
2803 * Attribute Swizzle Enable is set, Number of Output Attributes-1 if
2804 * enable is not set.
2805 * read_length = ceiling((max_source_attr + 1) / 2)
2807 * [errata] Corruption/Hang possible if length programmed larger than
2810 * Similar text exists for Ivy Bridge.
2812 * We find the last URB slot that's actually read by the FS.
2814 unsigned last_read_slot
= last_vue_map
->num_slots
- 1;
2815 while (last_read_slot
> first_slot
&& !(fs_input_slots
&
2816 (1ull << last_vue_map
->slot_to_varying
[last_read_slot
])))
2819 /* The URB read length is the difference of the two, counted in pairs. */
2820 *out_length
= DIV_ROUND_UP(last_read_slot
- first_slot
+ 1, 2);
2824 iris_emit_sbe_swiz(struct iris_batch
*batch
,
2825 const struct iris_context
*ice
,
2826 unsigned urb_read_offset
,
2827 unsigned sprite_coord_enables
)
2829 struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL
) attr_overrides
[16] = {};
2830 const struct brw_wm_prog_data
*wm_prog_data
= (void *)
2831 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
2832 const struct brw_vue_map
*vue_map
= ice
->shaders
.last_vue_map
;
2833 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
2835 /* XXX: this should be generated when putting programs in place */
2837 // XXX: raster->sprite_coord_enable
2839 for (int fs_attr
= 0; fs_attr
< VARYING_SLOT_MAX
; fs_attr
++) {
2840 const int input_index
= wm_prog_data
->urb_setup
[fs_attr
];
2841 if (input_index
< 0 || input_index
>= 16)
2844 struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL
) *attr
=
2845 &attr_overrides
[input_index
];
2846 int slot
= vue_map
->varying_to_slot
[fs_attr
];
2848 /* Viewport and Layer are stored in the VUE header. We need to override
2849 * them to zero if earlier stages didn't write them, as GL requires that
2850 * they read back as zero when not explicitly set.
2853 case VARYING_SLOT_VIEWPORT
:
2854 case VARYING_SLOT_LAYER
:
2855 attr
->ComponentOverrideX
= true;
2856 attr
->ComponentOverrideW
= true;
2857 attr
->ConstantSource
= CONST_0000
;
2859 if (!(vue_map
->slots_valid
& VARYING_BIT_LAYER
))
2860 attr
->ComponentOverrideY
= true;
2861 if (!(vue_map
->slots_valid
& VARYING_BIT_VIEWPORT
))
2862 attr
->ComponentOverrideZ
= true;
2865 case VARYING_SLOT_PRIMITIVE_ID
:
2866 /* Override if the previous shader stage didn't write gl_PrimitiveID. */
2868 attr
->ComponentOverrideX
= true;
2869 attr
->ComponentOverrideY
= true;
2870 attr
->ComponentOverrideZ
= true;
2871 attr
->ComponentOverrideW
= true;
2872 attr
->ConstantSource
= PRIM_ID
;
2880 if (sprite_coord_enables
& (1 << input_index
))
2883 /* If there was only a back color written but not front, use back
2884 * as the color instead of undefined.
2886 if (slot
== -1 && fs_attr
== VARYING_SLOT_COL0
)
2887 slot
= vue_map
->varying_to_slot
[VARYING_SLOT_BFC0
];
2888 if (slot
== -1 && fs_attr
== VARYING_SLOT_COL1
)
2889 slot
= vue_map
->varying_to_slot
[VARYING_SLOT_BFC1
];
2891 /* Not written by the previous stage - undefined. */
2893 attr
->ComponentOverrideX
= true;
2894 attr
->ComponentOverrideY
= true;
2895 attr
->ComponentOverrideZ
= true;
2896 attr
->ComponentOverrideW
= true;
2897 attr
->ConstantSource
= CONST_0001_FLOAT
;
2901 /* Compute the location of the attribute relative to the read offset,
2902 * which is counted in 256-bit increments (two 128-bit VUE slots).
2904 const int source_attr
= slot
- 2 * urb_read_offset
;
2905 assert(source_attr
>= 0 && source_attr
<= 32);
2906 attr
->SourceAttribute
= source_attr
;
2908 /* If we are doing two-sided color, and the VUE slot following this one
2909 * represents a back-facing color, then we need to instruct the SF unit
2910 * to do back-facing swizzling.
2912 if (cso_rast
->light_twoside
&&
2913 ((vue_map
->slot_to_varying
[slot
] == VARYING_SLOT_COL0
&&
2914 vue_map
->slot_to_varying
[slot
+1] == VARYING_SLOT_BFC0
) ||
2915 (vue_map
->slot_to_varying
[slot
] == VARYING_SLOT_COL1
&&
2916 vue_map
->slot_to_varying
[slot
+1] == VARYING_SLOT_BFC1
)))
2917 attr
->SwizzleSelect
= INPUTATTR_FACING
;
2920 iris_emit_cmd(batch
, GENX(3DSTATE_SBE_SWIZ
), sbes
) {
2921 for (int i
= 0; i
< 16; i
++)
2922 sbes
.Attribute
[i
] = attr_overrides
[i
];
2927 iris_calculate_point_sprite_overrides(const struct brw_wm_prog_data
*prog_data
,
2928 const struct iris_rasterizer_state
*cso
)
2930 unsigned overrides
= 0;
2932 if (prog_data
->urb_setup
[VARYING_SLOT_PNTC
] != -1)
2933 overrides
|= 1 << prog_data
->urb_setup
[VARYING_SLOT_PNTC
];
2935 for (int i
= 0; i
< 8; i
++) {
2936 if ((cso
->sprite_coord_enable
& (1 << i
)) &&
2937 prog_data
->urb_setup
[VARYING_SLOT_TEX0
+ i
] != -1)
2938 overrides
|= 1 << prog_data
->urb_setup
[VARYING_SLOT_TEX0
+ i
];
2945 iris_emit_sbe(struct iris_batch
*batch
, const struct iris_context
*ice
)
2947 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
2948 const struct brw_wm_prog_data
*wm_prog_data
= (void *)
2949 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
2950 const struct shader_info
*fs_info
=
2951 iris_get_shader_info(ice
, MESA_SHADER_FRAGMENT
);
2953 unsigned urb_read_offset
, urb_read_length
;
2954 iris_compute_sbe_urb_read_interval(fs_info
->inputs_read
,
2955 ice
->shaders
.last_vue_map
,
2956 cso_rast
->light_twoside
,
2957 &urb_read_offset
, &urb_read_length
);
2959 unsigned sprite_coord_overrides
=
2960 iris_calculate_point_sprite_overrides(wm_prog_data
, cso_rast
);
2962 iris_emit_cmd(batch
, GENX(3DSTATE_SBE
), sbe
) {
2963 sbe
.AttributeSwizzleEnable
= true;
2964 sbe
.NumberofSFOutputAttributes
= wm_prog_data
->num_varying_inputs
;
2965 sbe
.PointSpriteTextureCoordinateOrigin
= cso_rast
->sprite_coord_mode
;
2966 sbe
.VertexURBEntryReadOffset
= urb_read_offset
;
2967 sbe
.VertexURBEntryReadLength
= urb_read_length
;
2968 sbe
.ForceVertexURBEntryReadOffset
= true;
2969 sbe
.ForceVertexURBEntryReadLength
= true;
2970 sbe
.ConstantInterpolationEnable
= wm_prog_data
->flat_inputs
;
2971 sbe
.PointSpriteTextureCoordinateEnable
= sprite_coord_overrides
;
2973 for (int i
= 0; i
< 32; i
++) {
2974 sbe
.AttributeActiveComponentFormat
[i
] = ACTIVE_COMPONENT_XYZW
;
2978 iris_emit_sbe_swiz(batch
, ice
, urb_read_offset
, sprite_coord_overrides
);
2981 /* ------------------------------------------------------------------- */
2984 * Populate VS program key fields based on the current state.
2987 iris_populate_vs_key(const struct iris_context
*ice
,
2988 const struct shader_info
*info
,
2989 struct brw_vs_prog_key
*key
)
2991 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
2993 if (info
->clip_distance_array_size
== 0 &&
2994 (info
->outputs_written
& (VARYING_BIT_POS
| VARYING_BIT_CLIP_VERTEX
)))
2995 key
->nr_userclip_plane_consts
= cso_rast
->num_clip_plane_consts
;
2999 * Populate TCS program key fields based on the current state.
3002 iris_populate_tcs_key(const struct iris_context
*ice
,
3003 struct brw_tcs_prog_key
*key
)
3008 * Populate TES program key fields based on the current state.
3011 iris_populate_tes_key(const struct iris_context
*ice
,
3012 struct brw_tes_prog_key
*key
)
3017 * Populate GS program key fields based on the current state.
3020 iris_populate_gs_key(const struct iris_context
*ice
,
3021 struct brw_gs_prog_key
*key
)
3026 * Populate FS program key fields based on the current state.
3029 iris_populate_fs_key(const struct iris_context
*ice
,
3030 struct brw_wm_prog_key
*key
)
3032 /* XXX: dirty flags? */
3033 const struct pipe_framebuffer_state
*fb
= &ice
->state
.framebuffer
;
3034 const struct iris_depth_stencil_alpha_state
*zsa
= ice
->state
.cso_zsa
;
3035 const struct iris_rasterizer_state
*rast
= ice
->state
.cso_rast
;
3036 const struct iris_blend_state
*blend
= ice
->state
.cso_blend
;
3038 key
->nr_color_regions
= fb
->nr_cbufs
;
3040 key
->clamp_fragment_color
= rast
->clamp_fragment_color
;
3042 key
->replicate_alpha
= fb
->nr_cbufs
> 1 &&
3043 (zsa
->alpha
.enabled
|| blend
->alpha_to_coverage
);
3045 /* XXX: only bother if COL0/1 are read */
3046 key
->flat_shade
= rast
->flatshade
;
3048 key
->persample_interp
= rast
->force_persample_interp
;
3049 key
->multisample_fbo
= rast
->multisample
&& fb
->samples
> 1;
3051 key
->coherent_fb_fetch
= true;
3053 // XXX: uint64_t input_slots_valid; - for >16 inputs
3055 // XXX: key->force_dual_color_blend for unigine
3056 // XXX: respect hint for high_quality_derivatives:1;
3060 iris_populate_cs_key(const struct iris_context
*ice
,
3061 struct brw_cs_prog_key
*key
)
3066 // XXX: these need to go in INIT_THREAD_DISPATCH_FIELDS
3067 pkt
.SamplerCount
= \
3068 DIV_ROUND_UP(CLAMP(stage_state
->sampler_count
, 0, 16), 4); \
3073 KSP(const struct iris_compiled_shader
*shader
)
3075 struct iris_resource
*res
= (void *) shader
->assembly
.res
;
3076 return iris_bo_offset_from_base_address(res
->bo
) + shader
->assembly
.offset
;
3079 // Gen11 workaround table #2056 WABTPPrefetchDisable suggests to disable
3080 // prefetching of binding tables in A0 and B0 steppings. XXX: Revisit
3081 // this WA on C0 stepping.
3083 #define INIT_THREAD_DISPATCH_FIELDS(pkt, prefix, stage) \
3084 pkt.KernelStartPointer = KSP(shader); \
3085 pkt.BindingTableEntryCount = GEN_GEN == 11 ? 0 : \
3086 prog_data->binding_table.size_bytes / 4; \
3087 pkt.FloatingPointMode = prog_data->use_alt_mode; \
3089 pkt.DispatchGRFStartRegisterForURBData = \
3090 prog_data->dispatch_grf_start_reg; \
3091 pkt.prefix##URBEntryReadLength = vue_prog_data->urb_read_length; \
3092 pkt.prefix##URBEntryReadOffset = 0; \
3094 pkt.StatisticsEnable = true; \
3095 pkt.Enable = true; \
3097 if (prog_data->total_scratch) { \
3098 uint32_t scratch_addr = \
3099 iris_get_scratch_space(ice, prog_data->total_scratch, stage); \
3100 pkt.PerThreadScratchSpace = ffs(prog_data->total_scratch) - 11; \
3101 pkt.ScratchSpaceBasePointer = rw_bo(NULL, scratch_addr); \
3105 * Encode most of 3DSTATE_VS based on the compiled shader.
3108 iris_store_vs_state(struct iris_context
*ice
,
3109 const struct gen_device_info
*devinfo
,
3110 struct iris_compiled_shader
*shader
)
3112 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3113 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3115 iris_pack_command(GENX(3DSTATE_VS
), shader
->derived_data
, vs
) {
3116 INIT_THREAD_DISPATCH_FIELDS(vs
, Vertex
, MESA_SHADER_VERTEX
);
3117 vs
.MaximumNumberofThreads
= devinfo
->max_vs_threads
- 1;
3118 vs
.SIMD8DispatchEnable
= true;
3119 vs
.UserClipDistanceCullTestEnableBitmask
=
3120 vue_prog_data
->cull_distance_mask
;
3125 * Encode most of 3DSTATE_HS based on the compiled shader.
3128 iris_store_tcs_state(struct iris_context
*ice
,
3129 const struct gen_device_info
*devinfo
,
3130 struct iris_compiled_shader
*shader
)
3132 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3133 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3134 struct brw_tcs_prog_data
*tcs_prog_data
= (void *) prog_data
;
3136 iris_pack_command(GENX(3DSTATE_HS
), shader
->derived_data
, hs
) {
3137 INIT_THREAD_DISPATCH_FIELDS(hs
, Vertex
, MESA_SHADER_TESS_CTRL
);
3139 hs
.InstanceCount
= tcs_prog_data
->instances
- 1;
3140 hs
.MaximumNumberofThreads
= devinfo
->max_tcs_threads
- 1;
3141 hs
.IncludeVertexHandles
= true;
3146 * Encode 3DSTATE_TE and most of 3DSTATE_DS based on the compiled shader.
3149 iris_store_tes_state(struct iris_context
*ice
,
3150 const struct gen_device_info
*devinfo
,
3151 struct iris_compiled_shader
*shader
)
3153 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3154 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3155 struct brw_tes_prog_data
*tes_prog_data
= (void *) prog_data
;
3157 uint32_t *te_state
= (void *) shader
->derived_data
;
3158 uint32_t *ds_state
= te_state
+ GENX(3DSTATE_TE_length
);
3160 iris_pack_command(GENX(3DSTATE_TE
), te_state
, te
) {
3161 te
.Partitioning
= tes_prog_data
->partitioning
;
3162 te
.OutputTopology
= tes_prog_data
->output_topology
;
3163 te
.TEDomain
= tes_prog_data
->domain
;
3165 te
.MaximumTessellationFactorOdd
= 63.0;
3166 te
.MaximumTessellationFactorNotOdd
= 64.0;
3169 iris_pack_command(GENX(3DSTATE_DS
), ds_state
, ds
) {
3170 INIT_THREAD_DISPATCH_FIELDS(ds
, Patch
, MESA_SHADER_TESS_EVAL
);
3172 ds
.DispatchMode
= DISPATCH_MODE_SIMD8_SINGLE_PATCH
;
3173 ds
.MaximumNumberofThreads
= devinfo
->max_tes_threads
- 1;
3174 ds
.ComputeWCoordinateEnable
=
3175 tes_prog_data
->domain
== BRW_TESS_DOMAIN_TRI
;
3177 ds
.UserClipDistanceCullTestEnableBitmask
=
3178 vue_prog_data
->cull_distance_mask
;
3184 * Encode most of 3DSTATE_GS based on the compiled shader.
3187 iris_store_gs_state(struct iris_context
*ice
,
3188 const struct gen_device_info
*devinfo
,
3189 struct iris_compiled_shader
*shader
)
3191 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3192 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3193 struct brw_gs_prog_data
*gs_prog_data
= (void *) prog_data
;
3195 iris_pack_command(GENX(3DSTATE_GS
), shader
->derived_data
, gs
) {
3196 INIT_THREAD_DISPATCH_FIELDS(gs
, Vertex
, MESA_SHADER_GEOMETRY
);
3198 gs
.OutputVertexSize
= gs_prog_data
->output_vertex_size_hwords
* 2 - 1;
3199 gs
.OutputTopology
= gs_prog_data
->output_topology
;
3200 gs
.ControlDataHeaderSize
=
3201 gs_prog_data
->control_data_header_size_hwords
;
3202 gs
.InstanceControl
= gs_prog_data
->invocations
- 1;
3203 gs
.DispatchMode
= DISPATCH_MODE_SIMD8
;
3204 gs
.IncludePrimitiveID
= gs_prog_data
->include_primitive_id
;
3205 gs
.ControlDataFormat
= gs_prog_data
->control_data_format
;
3206 gs
.ReorderMode
= TRAILING
;
3207 gs
.ExpectedVertexCount
= gs_prog_data
->vertices_in
;
3208 gs
.MaximumNumberofThreads
=
3209 GEN_GEN
== 8 ? (devinfo
->max_gs_threads
/ 2 - 1)
3210 : (devinfo
->max_gs_threads
- 1);
3212 if (gs_prog_data
->static_vertex_count
!= -1) {
3213 gs
.StaticOutput
= true;
3214 gs
.StaticOutputVertexCount
= gs_prog_data
->static_vertex_count
;
3216 gs
.IncludeVertexHandles
= vue_prog_data
->include_vue_handles
;
3218 gs
.UserClipDistanceCullTestEnableBitmask
=
3219 vue_prog_data
->cull_distance_mask
;
3221 const int urb_entry_write_offset
= 1;
3222 const uint32_t urb_entry_output_length
=
3223 DIV_ROUND_UP(vue_prog_data
->vue_map
.num_slots
, 2) -
3224 urb_entry_write_offset
;
3226 gs
.VertexURBEntryOutputReadOffset
= urb_entry_write_offset
;
3227 gs
.VertexURBEntryOutputLength
= MAX2(urb_entry_output_length
, 1);
3232 * Encode most of 3DSTATE_PS and 3DSTATE_PS_EXTRA based on the shader.
3235 iris_store_fs_state(struct iris_context
*ice
,
3236 const struct gen_device_info
*devinfo
,
3237 struct iris_compiled_shader
*shader
)
3239 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3240 struct brw_wm_prog_data
*wm_prog_data
= (void *) shader
->prog_data
;
3242 uint32_t *ps_state
= (void *) shader
->derived_data
;
3243 uint32_t *psx_state
= ps_state
+ GENX(3DSTATE_PS_length
);
3245 iris_pack_command(GENX(3DSTATE_PS
), ps_state
, ps
) {
3246 ps
.VectorMaskEnable
= true;
3247 //ps.SamplerCount = ...
3248 // XXX: WABTPPrefetchDisable, see above, drop at C0
3249 ps
.BindingTableEntryCount
= GEN_GEN
== 11 ? 0 :
3250 prog_data
->binding_table
.size_bytes
/ 4;
3251 ps
.FloatingPointMode
= prog_data
->use_alt_mode
;
3252 ps
.MaximumNumberofThreadsPerPSD
= 64 - (GEN_GEN
== 8 ? 2 : 1);
3254 ps
.PushConstantEnable
= shader
->num_system_values
> 0 ||
3255 prog_data
->ubo_ranges
[0].length
> 0;
3257 /* From the documentation for this packet:
3258 * "If the PS kernel does not need the Position XY Offsets to
3259 * compute a Position Value, then this field should be programmed
3260 * to POSOFFSET_NONE."
3262 * "SW Recommendation: If the PS kernel needs the Position Offsets
3263 * to compute a Position XY value, this field should match Position
3264 * ZW Interpolation Mode to ensure a consistent position.xyzw
3267 * We only require XY sample offsets. So, this recommendation doesn't
3268 * look useful at the moment. We might need this in future.
3270 ps
.PositionXYOffsetSelect
=
3271 wm_prog_data
->uses_pos_offset
? POSOFFSET_SAMPLE
: POSOFFSET_NONE
;
3272 ps
._8PixelDispatchEnable
= wm_prog_data
->dispatch_8
;
3273 ps
._16PixelDispatchEnable
= wm_prog_data
->dispatch_16
;
3274 ps
._32PixelDispatchEnable
= wm_prog_data
->dispatch_32
;
3276 // XXX: Disable SIMD32 with 16x MSAA
3278 ps
.DispatchGRFStartRegisterForConstantSetupData0
=
3279 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 0);
3280 ps
.DispatchGRFStartRegisterForConstantSetupData1
=
3281 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 1);
3282 ps
.DispatchGRFStartRegisterForConstantSetupData2
=
3283 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 2);
3285 ps
.KernelStartPointer0
=
3286 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 0);
3287 ps
.KernelStartPointer1
=
3288 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 1);
3289 ps
.KernelStartPointer2
=
3290 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 2);
3292 if (prog_data
->total_scratch
) {
3293 uint32_t scratch_addr
=
3294 iris_get_scratch_space(ice
, prog_data
->total_scratch
,
3295 MESA_SHADER_FRAGMENT
);
3296 ps
.PerThreadScratchSpace
= ffs(prog_data
->total_scratch
) - 11;
3297 ps
.ScratchSpaceBasePointer
= rw_bo(NULL
, scratch_addr
);
3301 iris_pack_command(GENX(3DSTATE_PS_EXTRA
), psx_state
, psx
) {
3302 psx
.PixelShaderValid
= true;
3303 psx
.PixelShaderComputedDepthMode
= wm_prog_data
->computed_depth_mode
;
3304 // XXX: alpha test / alpha to coverage :/
3305 psx
.PixelShaderKillsPixel
= wm_prog_data
->uses_kill
||
3306 wm_prog_data
->uses_omask
;
3307 psx
.AttributeEnable
= wm_prog_data
->num_varying_inputs
!= 0;
3308 psx
.PixelShaderUsesSourceDepth
= wm_prog_data
->uses_src_depth
;
3309 psx
.PixelShaderUsesSourceW
= wm_prog_data
->uses_src_w
;
3310 psx
.PixelShaderIsPerSample
= wm_prog_data
->persample_dispatch
;
3312 if (wm_prog_data
->uses_sample_mask
) {
3313 /* TODO: conservative rasterization */
3314 if (wm_prog_data
->post_depth_coverage
)
3315 psx
.InputCoverageMaskState
= ICMS_DEPTH_COVERAGE
;
3317 psx
.InputCoverageMaskState
= ICMS_NORMAL
;
3320 psx
.oMaskPresenttoRenderTarget
= wm_prog_data
->uses_omask
;
3321 psx
.PixelShaderPullsBary
= wm_prog_data
->pulls_bary
;
3322 psx
.PixelShaderComputesStencil
= wm_prog_data
->computed_stencil
;
3329 * Compute the size of the derived data (shader command packets).
3331 * This must match the data written by the iris_store_xs_state() functions.
3334 iris_store_cs_state(struct iris_context
*ice
,
3335 const struct gen_device_info
*devinfo
,
3336 struct iris_compiled_shader
*shader
)
3338 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3339 struct brw_cs_prog_data
*cs_prog_data
= (void *) shader
->prog_data
;
3340 void *map
= shader
->derived_data
;
3342 iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA
), map
, desc
) {
3343 desc
.KernelStartPointer
= KSP(shader
);
3344 desc
.ConstantURBEntryReadLength
= cs_prog_data
->push
.per_thread
.regs
;
3345 desc
.NumberofThreadsinGPGPUThreadGroup
= cs_prog_data
->threads
;
3346 desc
.SharedLocalMemorySize
=
3347 encode_slm_size(GEN_GEN
, prog_data
->total_shared
);
3348 desc
.BarrierEnable
= cs_prog_data
->uses_barrier
;
3349 desc
.CrossThreadConstantDataReadLength
=
3350 cs_prog_data
->push
.cross_thread
.regs
;
3355 iris_derived_program_state_size(enum iris_program_cache_id cache_id
)
3357 assert(cache_id
<= IRIS_CACHE_BLORP
);
3359 static const unsigned dwords
[] = {
3360 [IRIS_CACHE_VS
] = GENX(3DSTATE_VS_length
),
3361 [IRIS_CACHE_TCS
] = GENX(3DSTATE_HS_length
),
3362 [IRIS_CACHE_TES
] = GENX(3DSTATE_TE_length
) + GENX(3DSTATE_DS_length
),
3363 [IRIS_CACHE_GS
] = GENX(3DSTATE_GS_length
),
3365 GENX(3DSTATE_PS_length
) + GENX(3DSTATE_PS_EXTRA_length
),
3366 [IRIS_CACHE_CS
] = GENX(INTERFACE_DESCRIPTOR_DATA_length
),
3367 [IRIS_CACHE_BLORP
] = 0,
3370 return sizeof(uint32_t) * dwords
[cache_id
];
3374 * Create any state packets corresponding to the given shader stage
3375 * (i.e. 3DSTATE_VS) and save them as "derived data" in the shader variant.
3376 * This means that we can look up a program in the in-memory cache and
3377 * get most of the state packet without having to reconstruct it.
3380 iris_store_derived_program_state(struct iris_context
*ice
,
3381 enum iris_program_cache_id cache_id
,
3382 struct iris_compiled_shader
*shader
)
3384 struct iris_screen
*screen
= (void *) ice
->ctx
.screen
;
3385 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
3389 iris_store_vs_state(ice
, devinfo
, shader
);
3391 case IRIS_CACHE_TCS
:
3392 iris_store_tcs_state(ice
, devinfo
, shader
);
3394 case IRIS_CACHE_TES
:
3395 iris_store_tes_state(ice
, devinfo
, shader
);
3398 iris_store_gs_state(ice
, devinfo
, shader
);
3401 iris_store_fs_state(ice
, devinfo
, shader
);
3404 iris_store_cs_state(ice
, devinfo
, shader
);
3405 case IRIS_CACHE_BLORP
:
3412 /* ------------------------------------------------------------------- */
3415 * Configure the URB.
3417 * XXX: write a real comment.
3420 iris_upload_urb_config(struct iris_context
*ice
, struct iris_batch
*batch
)
3422 const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
3423 const unsigned push_size_kB
= 32;
3424 unsigned entries
[4];
3428 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
3429 if (!ice
->shaders
.prog
[i
]) {
3432 struct brw_vue_prog_data
*vue_prog_data
=
3433 (void *) ice
->shaders
.prog
[i
]->prog_data
;
3434 size
[i
] = vue_prog_data
->urb_entry_size
;
3436 assert(size
[i
] != 0);
3439 gen_get_urb_config(devinfo
, 1024 * push_size_kB
,
3440 1024 * ice
->shaders
.urb_size
,
3441 ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
] != NULL
,
3442 ice
->shaders
.prog
[MESA_SHADER_GEOMETRY
] != NULL
,
3443 size
, entries
, start
);
3445 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
3446 iris_emit_cmd(batch
, GENX(3DSTATE_URB_VS
), urb
) {
3447 urb
._3DCommandSubOpcode
+= i
;
3448 urb
.VSURBStartingAddress
= start
[i
];
3449 urb
.VSURBEntryAllocationSize
= size
[i
] - 1;
3450 urb
.VSNumberofURBEntries
= entries
[i
];
3455 static const uint32_t push_constant_opcodes
[] = {
3456 [MESA_SHADER_VERTEX
] = 21,
3457 [MESA_SHADER_TESS_CTRL
] = 25, /* HS */
3458 [MESA_SHADER_TESS_EVAL
] = 26, /* DS */
3459 [MESA_SHADER_GEOMETRY
] = 22,
3460 [MESA_SHADER_FRAGMENT
] = 23,
3461 [MESA_SHADER_COMPUTE
] = 0,
3465 use_null_surface(struct iris_batch
*batch
, struct iris_context
*ice
)
3467 struct iris_bo
*state_bo
= iris_resource_bo(ice
->state
.unbound_tex
.res
);
3469 iris_use_pinned_bo(batch
, state_bo
, false);
3471 return ice
->state
.unbound_tex
.offset
;
3475 use_null_fb_surface(struct iris_batch
*batch
, struct iris_context
*ice
)
3477 /* If set_framebuffer_state() was never called, fall back to 1x1x1 */
3478 if (!ice
->state
.null_fb
.res
)
3479 return use_null_surface(batch
, ice
);
3481 struct iris_bo
*state_bo
= iris_resource_bo(ice
->state
.null_fb
.res
);
3483 iris_use_pinned_bo(batch
, state_bo
, false);
3485 return ice
->state
.null_fb
.offset
;
3489 * Add a surface to the validation list, as well as the buffer containing
3490 * the corresponding SURFACE_STATE.
3492 * Returns the binding table entry (offset to SURFACE_STATE).
3495 use_surface(struct iris_batch
*batch
,
3496 struct pipe_surface
*p_surf
,
3499 struct iris_surface
*surf
= (void *) p_surf
;
3501 iris_use_pinned_bo(batch
, iris_resource_bo(p_surf
->texture
), writeable
);
3502 iris_use_pinned_bo(batch
, iris_resource_bo(surf
->surface_state
.res
), false);
3504 return surf
->surface_state
.offset
;
3508 use_sampler_view(struct iris_batch
*batch
, struct iris_sampler_view
*isv
)
3510 iris_use_pinned_bo(batch
, isv
->res
->bo
, false);
3511 iris_use_pinned_bo(batch
, iris_resource_bo(isv
->surface_state
.res
), false);
3513 return isv
->surface_state
.offset
;
3517 use_const_buffer(struct iris_batch
*batch
,
3518 struct iris_context
*ice
,
3519 struct iris_const_buffer
*cbuf
)
3521 if (!cbuf
->surface_state
.res
)
3522 return use_null_surface(batch
, ice
);
3524 iris_use_pinned_bo(batch
, iris_resource_bo(cbuf
->data
.res
), false);
3525 iris_use_pinned_bo(batch
, iris_resource_bo(cbuf
->surface_state
.res
), false);
3527 return cbuf
->surface_state
.offset
;
3531 use_ssbo(struct iris_batch
*batch
, struct iris_context
*ice
,
3532 struct iris_shader_state
*shs
, int i
)
3535 return use_null_surface(batch
, ice
);
3537 struct iris_state_ref
*surf_state
= &shs
->ssbo_surface_state
[i
];
3539 iris_use_pinned_bo(batch
, iris_resource_bo(shs
->ssbo
[i
]), true);
3540 iris_use_pinned_bo(batch
, iris_resource_bo(surf_state
->res
), false);
3542 return surf_state
->offset
;
3546 use_image(struct iris_batch
*batch
, struct iris_context
*ice
,
3547 struct iris_shader_state
*shs
, int i
)
3549 if (!shs
->image
[i
].res
)
3550 return use_null_surface(batch
, ice
);
3552 struct iris_state_ref
*surf_state
= &shs
->image
[i
].surface_state
;
3554 iris_use_pinned_bo(batch
, iris_resource_bo(shs
->image
[i
].res
),
3555 shs
->image
[i
].access
& PIPE_IMAGE_ACCESS_WRITE
);
3556 iris_use_pinned_bo(batch
, iris_resource_bo(surf_state
->res
), false);
3558 return surf_state
->offset
;
3561 #define push_bt_entry(addr) \
3562 assert(addr >= binder_addr); \
3563 assert(s < prog_data->binding_table.size_bytes / sizeof(uint32_t)); \
3564 if (!pin_only) bt_map[s++] = (addr) - binder_addr;
3566 #define bt_assert(section, exists) \
3567 if (!pin_only) assert(prog_data->binding_table.section == \
3568 (exists) ? s : 0xd0d0d0d0)
3571 * Populate the binding table for a given shader stage.
3573 * This fills out the table of pointers to surfaces required by the shader,
3574 * and also adds those buffers to the validation list so the kernel can make
3575 * resident before running our batch.
3578 iris_populate_binding_table(struct iris_context
*ice
,
3579 struct iris_batch
*batch
,
3580 gl_shader_stage stage
,
3583 const struct iris_binder
*binder
= &ice
->state
.binder
;
3584 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3588 UNUSED
struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3589 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3590 uint32_t binder_addr
= binder
->bo
->gtt_offset
;
3592 //struct brw_stage_prog_data *prog_data = (void *) shader->prog_data;
3593 uint32_t *bt_map
= binder
->map
+ binder
->bt_offset
[stage
];
3596 const struct shader_info
*info
= iris_get_shader_info(ice
, stage
);
3598 /* TCS passthrough doesn't need a binding table. */
3599 assert(stage
== MESA_SHADER_TESS_CTRL
);
3603 if (stage
== MESA_SHADER_COMPUTE
) {
3604 /* surface for gl_NumWorkGroups */
3605 struct iris_state_ref
*grid_data
= &ice
->state
.grid_size
;
3606 struct iris_state_ref
*grid_state
= &ice
->state
.grid_surf_state
;
3607 iris_use_pinned_bo(batch
, iris_resource_bo(grid_data
->res
), false);
3608 iris_use_pinned_bo(batch
, iris_resource_bo(grid_state
->res
), false);
3609 push_bt_entry(grid_state
->offset
);
3612 if (stage
== MESA_SHADER_FRAGMENT
) {
3613 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
3614 /* Note that cso_fb->nr_cbufs == fs_key->nr_color_regions. */
3615 if (cso_fb
->nr_cbufs
) {
3616 for (unsigned i
= 0; i
< cso_fb
->nr_cbufs
; i
++) {
3618 cso_fb
->cbufs
[i
] ? use_surface(batch
, cso_fb
->cbufs
[i
], true)
3619 : use_null_fb_surface(batch
, ice
);
3620 push_bt_entry(addr
);
3623 uint32_t addr
= use_null_fb_surface(batch
, ice
);
3624 push_bt_entry(addr
);
3628 bt_assert(texture_start
, info
->num_textures
> 0);
3630 for (int i
= 0; i
< info
->num_textures
; i
++) {
3631 struct iris_sampler_view
*view
= shs
->textures
[i
];
3632 uint32_t addr
= view
? use_sampler_view(batch
, view
)
3633 : use_null_surface(batch
, ice
);
3634 push_bt_entry(addr
);
3637 bt_assert(image_start
, info
->num_images
> 0);
3639 for (int i
= 0; i
< info
->num_images
; i
++) {
3640 uint32_t addr
= use_image(batch
, ice
, shs
, i
);
3641 push_bt_entry(addr
);
3644 const int num_ubos
= iris_get_shader_num_ubos(ice
, stage
);
3646 bt_assert(ubo_start
, num_ubos
> 0);
3648 for (int i
= 0; i
< num_ubos
; i
++) {
3649 uint32_t addr
= use_const_buffer(batch
, ice
, &shs
->constbuf
[i
]);
3650 push_bt_entry(addr
);
3653 bt_assert(ssbo_start
, info
->num_abos
+ info
->num_ssbos
> 0);
3655 /* XXX: st is wasting 16 binding table slots for ABOs. Should add a cap
3656 * for changing nir_lower_atomics_to_ssbos setting and buffer_base offset
3657 * in st_atom_storagebuf.c so it'll compact them into one range, with
3658 * SSBOs starting at info->num_abos. Ideally it'd reset num_abos to 0 too
3660 if (info
->num_abos
+ info
->num_ssbos
> 0) {
3661 for (int i
= 0; i
< IRIS_MAX_ABOS
+ info
->num_ssbos
; i
++) {
3662 uint32_t addr
= use_ssbo(batch
, ice
, shs
, i
);
3663 push_bt_entry(addr
);
3668 // XXX: not implemented yet
3669 bt_assert(plane_start
[1], ...);
3670 bt_assert(plane_start
[2], ...);
3675 iris_use_optional_res(struct iris_batch
*batch
,
3676 struct pipe_resource
*res
,
3680 struct iris_bo
*bo
= iris_resource_bo(res
);
3681 iris_use_pinned_bo(batch
, bo
, writeable
);
3685 /* ------------------------------------------------------------------- */
3688 * Pin any BOs which were installed by a previous batch, and restored
3689 * via the hardware logical context mechanism.
3691 * We don't need to re-emit all state every batch - the hardware context
3692 * mechanism will save and restore it for us. This includes pointers to
3693 * various BOs...which won't exist unless we ask the kernel to pin them
3694 * by adding them to the validation list.
3696 * We can skip buffers if we've re-emitted those packets, as we're
3697 * overwriting those stale pointers with new ones, and don't actually
3698 * refer to the old BOs.
3701 iris_restore_render_saved_bos(struct iris_context
*ice
,
3702 struct iris_batch
*batch
,
3703 const struct pipe_draw_info
*draw
)
3705 struct iris_genx_state
*genx
= ice
->state
.genx
;
3707 // XXX: whack IRIS_SHADER_DIRTY_BINDING_TABLE on new batch
3709 const uint64_t clean
= ~ice
->state
.dirty
;
3711 if (clean
& IRIS_DIRTY_CC_VIEWPORT
) {
3712 iris_use_optional_res(batch
, ice
->state
.last_res
.cc_vp
, false);
3715 if (clean
& IRIS_DIRTY_SF_CL_VIEWPORT
) {
3716 iris_use_optional_res(batch
, ice
->state
.last_res
.sf_cl_vp
, false);
3719 if (clean
& IRIS_DIRTY_BLEND_STATE
) {
3720 iris_use_optional_res(batch
, ice
->state
.last_res
.blend
, false);
3723 if (clean
& IRIS_DIRTY_COLOR_CALC_STATE
) {
3724 iris_use_optional_res(batch
, ice
->state
.last_res
.color_calc
, false);
3727 if (clean
& IRIS_DIRTY_SCISSOR_RECT
) {
3728 iris_use_optional_res(batch
, ice
->state
.last_res
.scissor
, false);
3731 if (ice
->state
.streamout_active
&& (clean
& IRIS_DIRTY_SO_BUFFERS
)) {
3732 for (int i
= 0; i
< 4; i
++) {
3733 struct iris_stream_output_target
*tgt
=
3734 (void *) ice
->state
.so_target
[i
];
3736 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->base
.buffer
),
3738 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->offset
.res
),
3744 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3745 if (!(clean
& (IRIS_DIRTY_CONSTANTS_VS
<< stage
)))
3748 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3749 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3754 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
3756 for (int i
= 0; i
< 4; i
++) {
3757 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
3759 if (range
->length
== 0)
3762 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
3763 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
3766 iris_use_pinned_bo(batch
, res
->bo
, false);
3768 iris_use_pinned_bo(batch
, batch
->screen
->workaround_bo
, false);
3772 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3773 if (clean
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
3774 /* Re-pin any buffers referred to by the binding table. */
3775 iris_populate_binding_table(ice
, batch
, stage
, true);
3779 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3780 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3781 struct pipe_resource
*res
= shs
->sampler_table
.res
;
3783 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
3786 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3787 if (clean
& (IRIS_DIRTY_VS
<< stage
)) {
3788 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3790 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
3791 iris_use_pinned_bo(batch
, bo
, false);
3794 // XXX: scratch buffer
3798 if (clean
& IRIS_DIRTY_DEPTH_BUFFER
) {
3799 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
3801 if (cso_fb
->zsbuf
) {
3802 struct iris_resource
*zres
, *sres
;
3803 iris_get_depth_stencil_resources(cso_fb
->zsbuf
->texture
,
3805 // XXX: might not be writable...
3807 iris_use_pinned_bo(batch
, zres
->bo
, true);
3809 iris_use_pinned_bo(batch
, sres
->bo
, true);
3813 if (draw
->index_size
== 0 && ice
->state
.last_res
.index_buffer
) {
3814 /* This draw didn't emit a new index buffer, so we are inheriting the
3815 * older index buffer. This draw didn't need it, but future ones may.
3817 struct iris_bo
*bo
= iris_resource_bo(ice
->state
.last_res
.index_buffer
);
3818 iris_use_pinned_bo(batch
, bo
, false);
3821 if (clean
& IRIS_DIRTY_VERTEX_BUFFERS
) {
3822 uint64_t bound
= ice
->state
.bound_vertex_buffers
;
3824 const int i
= u_bit_scan64(&bound
);
3825 struct pipe_resource
*res
= genx
->vertex_buffers
[i
].resource
;
3826 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
3832 iris_restore_compute_saved_bos(struct iris_context
*ice
,
3833 struct iris_batch
*batch
,
3834 const struct pipe_grid_info
*grid
)
3836 const uint64_t clean
= ~ice
->state
.dirty
;
3838 const int stage
= MESA_SHADER_COMPUTE
;
3839 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3841 if (clean
& IRIS_DIRTY_CONSTANTS_CS
) {
3842 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3845 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
3846 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[0];
3848 if (range
->length
> 0) {
3849 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
3850 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
3853 iris_use_pinned_bo(batch
, res
->bo
, false);
3855 iris_use_pinned_bo(batch
, batch
->screen
->workaround_bo
, false);
3860 if (clean
& IRIS_DIRTY_BINDINGS_CS
) {
3861 /* Re-pin any buffers referred to by the binding table. */
3862 iris_populate_binding_table(ice
, batch
, stage
, true);
3865 struct pipe_resource
*sampler_res
= shs
->sampler_table
.res
;
3867 iris_use_pinned_bo(batch
, iris_resource_bo(sampler_res
), false);
3869 if (clean
& IRIS_DIRTY_CS
) {
3870 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3872 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
3873 iris_use_pinned_bo(batch
, bo
, false);
3876 // XXX: scratch buffer
3881 * Possibly emit STATE_BASE_ADDRESS to update Surface State Base Address.
3884 iris_update_surface_base_address(struct iris_batch
*batch
,
3885 struct iris_binder
*binder
)
3887 if (batch
->last_surface_base_address
== binder
->bo
->gtt_offset
)
3890 flush_for_state_base_change(batch
);
3892 iris_emit_cmd(batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
3893 // XXX: sba.SurfaceStateMemoryObjectControlState = MOCS_WB;
3894 sba
.SurfaceStateBaseAddressModifyEnable
= true;
3895 sba
.SurfaceStateBaseAddress
= ro_bo(binder
->bo
, 0);
3898 batch
->last_surface_base_address
= binder
->bo
->gtt_offset
;
3902 iris_upload_dirty_render_state(struct iris_context
*ice
,
3903 struct iris_batch
*batch
,
3904 const struct pipe_draw_info
*draw
)
3906 const uint64_t dirty
= ice
->state
.dirty
;
3908 if (!(dirty
& IRIS_ALL_DIRTY_FOR_RENDER
))
3911 struct iris_genx_state
*genx
= ice
->state
.genx
;
3912 struct iris_binder
*binder
= &ice
->state
.binder
;
3913 struct brw_wm_prog_data
*wm_prog_data
= (void *)
3914 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
3916 if (dirty
& IRIS_DIRTY_CC_VIEWPORT
) {
3917 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3918 uint32_t cc_vp_address
;
3920 /* XXX: could avoid streaming for depth_clip [0,1] case. */
3921 uint32_t *cc_vp_map
=
3922 stream_state(batch
, ice
->state
.dynamic_uploader
,
3923 &ice
->state
.last_res
.cc_vp
,
3924 4 * ice
->state
.num_viewports
*
3925 GENX(CC_VIEWPORT_length
), 32, &cc_vp_address
);
3926 for (int i
= 0; i
< ice
->state
.num_viewports
; i
++) {
3928 util_viewport_zmin_zmax(&ice
->state
.viewports
[i
],
3929 cso_rast
->clip_halfz
, &zmin
, &zmax
);
3930 if (cso_rast
->depth_clip_near
)
3932 if (cso_rast
->depth_clip_far
)
3935 iris_pack_state(GENX(CC_VIEWPORT
), cc_vp_map
, ccv
) {
3936 ccv
.MinimumDepth
= zmin
;
3937 ccv
.MaximumDepth
= zmax
;
3940 cc_vp_map
+= GENX(CC_VIEWPORT_length
);
3943 iris_emit_cmd(batch
, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_CC
), ptr
) {
3944 ptr
.CCViewportPointer
= cc_vp_address
;
3948 if (dirty
& IRIS_DIRTY_SF_CL_VIEWPORT
) {
3949 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
3950 uint32_t sf_cl_vp_address
;
3952 stream_state(batch
, ice
->state
.dynamic_uploader
,
3953 &ice
->state
.last_res
.sf_cl_vp
,
3954 4 * ice
->state
.num_viewports
*
3955 GENX(SF_CLIP_VIEWPORT_length
), 64, &sf_cl_vp_address
);
3957 for (unsigned i
= 0; i
< ice
->state
.num_viewports
; i
++) {
3958 const struct pipe_viewport_state
*state
= &ice
->state
.viewports
[i
];
3959 float gb_xmin
, gb_xmax
, gb_ymin
, gb_ymax
;
3961 float vp_xmin
= viewport_extent(state
, 0, -1.0f
);
3962 float vp_xmax
= viewport_extent(state
, 0, 1.0f
);
3963 float vp_ymin
= viewport_extent(state
, 1, -1.0f
);
3964 float vp_ymax
= viewport_extent(state
, 1, 1.0f
);
3966 calculate_guardband_size(cso_fb
->width
, cso_fb
->height
,
3967 state
->scale
[0], state
->scale
[1],
3968 state
->translate
[0], state
->translate
[1],
3969 &gb_xmin
, &gb_xmax
, &gb_ymin
, &gb_ymax
);
3971 iris_pack_state(GENX(SF_CLIP_VIEWPORT
), vp_map
, vp
) {
3972 vp
.ViewportMatrixElementm00
= state
->scale
[0];
3973 vp
.ViewportMatrixElementm11
= state
->scale
[1];
3974 vp
.ViewportMatrixElementm22
= state
->scale
[2];
3975 vp
.ViewportMatrixElementm30
= state
->translate
[0];
3976 vp
.ViewportMatrixElementm31
= state
->translate
[1];
3977 vp
.ViewportMatrixElementm32
= state
->translate
[2];
3978 vp
.XMinClipGuardband
= gb_xmin
;
3979 vp
.XMaxClipGuardband
= gb_xmax
;
3980 vp
.YMinClipGuardband
= gb_ymin
;
3981 vp
.YMaxClipGuardband
= gb_ymax
;
3982 vp
.XMinViewPort
= MAX2(vp_xmin
, 0);
3983 vp
.XMaxViewPort
= MIN2(vp_xmax
, cso_fb
->width
) - 1;
3984 vp
.YMinViewPort
= MAX2(vp_ymin
, 0);
3985 vp
.YMaxViewPort
= MIN2(vp_ymax
, cso_fb
->height
) - 1;
3988 vp_map
+= GENX(SF_CLIP_VIEWPORT_length
);
3991 iris_emit_cmd(batch
, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP
), ptr
) {
3992 ptr
.SFClipViewportPointer
= sf_cl_vp_address
;
3998 // XXX: this is only flagged at setup, we assume a static configuration
3999 if (dirty
& IRIS_DIRTY_URB
) {
4000 iris_upload_urb_config(ice
, batch
);
4003 if (dirty
& IRIS_DIRTY_BLEND_STATE
) {
4004 struct iris_blend_state
*cso_blend
= ice
->state
.cso_blend
;
4005 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4006 struct iris_depth_stencil_alpha_state
*cso_zsa
= ice
->state
.cso_zsa
;
4007 const int header_dwords
= GENX(BLEND_STATE_length
);
4008 const int rt_dwords
= cso_fb
->nr_cbufs
* GENX(BLEND_STATE_ENTRY_length
);
4009 uint32_t blend_offset
;
4010 uint32_t *blend_map
=
4011 stream_state(batch
, ice
->state
.dynamic_uploader
,
4012 &ice
->state
.last_res
.blend
,
4013 4 * (header_dwords
+ rt_dwords
), 64, &blend_offset
);
4015 uint32_t blend_state_header
;
4016 iris_pack_state(GENX(BLEND_STATE
), &blend_state_header
, bs
) {
4017 bs
.AlphaTestEnable
= cso_zsa
->alpha
.enabled
;
4018 bs
.AlphaTestFunction
= translate_compare_func(cso_zsa
->alpha
.func
);
4021 blend_map
[0] = blend_state_header
| cso_blend
->blend_state
[0];
4022 memcpy(&blend_map
[1], &cso_blend
->blend_state
[1], 4 * rt_dwords
);
4024 iris_emit_cmd(batch
, GENX(3DSTATE_BLEND_STATE_POINTERS
), ptr
) {
4025 ptr
.BlendStatePointer
= blend_offset
;
4026 ptr
.BlendStatePointerValid
= true;
4030 if (dirty
& IRIS_DIRTY_COLOR_CALC_STATE
) {
4031 struct iris_depth_stencil_alpha_state
*cso
= ice
->state
.cso_zsa
;
4034 stream_state(batch
, ice
->state
.dynamic_uploader
,
4035 &ice
->state
.last_res
.color_calc
,
4036 sizeof(uint32_t) * GENX(COLOR_CALC_STATE_length
),
4038 iris_pack_state(GENX(COLOR_CALC_STATE
), cc_map
, cc
) {
4039 cc
.AlphaTestFormat
= ALPHATEST_FLOAT32
;
4040 cc
.AlphaReferenceValueAsFLOAT32
= cso
->alpha
.ref_value
;
4041 cc
.BlendConstantColorRed
= ice
->state
.blend_color
.color
[0];
4042 cc
.BlendConstantColorGreen
= ice
->state
.blend_color
.color
[1];
4043 cc
.BlendConstantColorBlue
= ice
->state
.blend_color
.color
[2];
4044 cc
.BlendConstantColorAlpha
= ice
->state
.blend_color
.color
[3];
4046 iris_emit_cmd(batch
, GENX(3DSTATE_CC_STATE_POINTERS
), ptr
) {
4047 ptr
.ColorCalcStatePointer
= cc_offset
;
4048 ptr
.ColorCalcStatePointerValid
= true;
4052 /* Upload constants for TCS passthrough. */
4053 if ((dirty
& IRIS_DIRTY_CONSTANTS_TCS
) &&
4054 ice
->shaders
.prog
[MESA_SHADER_TESS_CTRL
] &&
4055 !ice
->shaders
.uncompiled
[MESA_SHADER_TESS_CTRL
]) {
4056 struct iris_compiled_shader
*tes_shader
= ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
];
4059 /* Passthrough always copies 2 vec4s, so when uploading data we ensure
4060 * it is in the right layout for TES.
4063 struct brw_tes_prog_data
*tes_prog_data
= (void *) tes_shader
->prog_data
;
4064 switch (tes_prog_data
->domain
) {
4065 case BRW_TESS_DOMAIN_QUAD
:
4066 for (int i
= 0; i
< 4; i
++)
4067 hdr
[7 - i
] = ice
->state
.default_outer_level
[i
];
4068 hdr
[3] = ice
->state
.default_inner_level
[0];
4069 hdr
[2] = ice
->state
.default_inner_level
[1];
4071 case BRW_TESS_DOMAIN_TRI
:
4072 for (int i
= 0; i
< 3; i
++)
4073 hdr
[7 - i
] = ice
->state
.default_outer_level
[i
];
4074 hdr
[4] = ice
->state
.default_inner_level
[0];
4076 case BRW_TESS_DOMAIN_ISOLINE
:
4077 hdr
[7] = ice
->state
.default_outer_level
[1];
4078 hdr
[6] = ice
->state
.default_outer_level
[0];
4082 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_TESS_CTRL
];
4083 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[0];
4084 u_upload_data(ice
->ctx
.const_uploader
, 0, sizeof(hdr
), 32,
4085 &hdr
[0], &cbuf
->data
.offset
,
4089 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4090 if (!(dirty
& (IRIS_DIRTY_CONSTANTS_VS
<< stage
)))
4093 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4094 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4099 if (shs
->cbuf0_needs_upload
)
4100 upload_uniforms(ice
, stage
);
4102 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
4104 iris_emit_cmd(batch
, GENX(3DSTATE_CONSTANT_VS
), pkt
) {
4105 pkt
._3DCommandSubOpcode
= push_constant_opcodes
[stage
];
4107 /* The Skylake PRM contains the following restriction:
4109 * "The driver must ensure The following case does not occur
4110 * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
4111 * buffer 3 read length equal to zero committed followed by a
4112 * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to
4115 * To avoid this, we program the buffers in the highest slots.
4116 * This way, slot 0 is only used if slot 3 is also used.
4120 for (int i
= 3; i
>= 0; i
--) {
4121 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
4123 if (range
->length
== 0)
4126 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
4127 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
4129 assert(cbuf
->data
.offset
% 32 == 0);
4131 pkt
.ConstantBody
.ReadLength
[n
] = range
->length
;
4132 pkt
.ConstantBody
.Buffer
[n
] =
4133 res
? ro_bo(res
->bo
, range
->start
* 32 + cbuf
->data
.offset
)
4134 : ro_bo(batch
->screen
->workaround_bo
, 0);
4141 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4142 if (dirty
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4143 iris_emit_cmd(batch
, GENX(3DSTATE_BINDING_TABLE_POINTERS_VS
), ptr
) {
4144 ptr
._3DCommandSubOpcode
= 38 + stage
;
4145 ptr
.PointertoVSBindingTable
= binder
->bt_offset
[stage
];
4150 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4151 if (dirty
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4152 iris_populate_binding_table(ice
, batch
, stage
, false);
4156 if (ice
->state
.need_border_colors
)
4157 iris_use_pinned_bo(batch
, ice
->state
.border_color_pool
.bo
, false);
4159 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4160 if (!(dirty
& (IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
)) ||
4161 !ice
->shaders
.prog
[stage
])
4164 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4165 struct pipe_resource
*res
= shs
->sampler_table
.res
;
4167 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
4169 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLER_STATE_POINTERS_VS
), ptr
) {
4170 ptr
._3DCommandSubOpcode
= 43 + stage
;
4171 ptr
.PointertoVSSamplerState
= shs
->sampler_table
.offset
;
4175 if (dirty
& IRIS_DIRTY_MULTISAMPLE
) {
4176 iris_emit_cmd(batch
, GENX(3DSTATE_MULTISAMPLE
), ms
) {
4178 ice
->state
.cso_rast
->half_pixel_center
? CENTER
: UL_CORNER
;
4179 if (ice
->state
.framebuffer
.samples
> 0)
4180 ms
.NumberofMultisamples
= ffs(ice
->state
.framebuffer
.samples
) - 1;
4184 if (dirty
& IRIS_DIRTY_SAMPLE_MASK
) {
4185 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLE_MASK
), ms
) {
4186 ms
.SampleMask
= MAX2(ice
->state
.sample_mask
, 1);
4190 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4191 if (!(dirty
& (IRIS_DIRTY_VS
<< stage
)))
4194 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4197 struct iris_resource
*cache
= (void *) shader
->assembly
.res
;
4198 iris_use_pinned_bo(batch
, cache
->bo
, false);
4199 iris_batch_emit(batch
, shader
->derived_data
,
4200 iris_derived_program_state_size(stage
));
4202 if (stage
== MESA_SHADER_TESS_EVAL
) {
4203 iris_emit_cmd(batch
, GENX(3DSTATE_HS
), hs
);
4204 iris_emit_cmd(batch
, GENX(3DSTATE_TE
), te
);
4205 iris_emit_cmd(batch
, GENX(3DSTATE_DS
), ds
);
4206 } else if (stage
== MESA_SHADER_GEOMETRY
) {
4207 iris_emit_cmd(batch
, GENX(3DSTATE_GS
), gs
);
4212 if (ice
->state
.streamout_active
) {
4213 if (dirty
& IRIS_DIRTY_SO_BUFFERS
) {
4214 iris_batch_emit(batch
, genx
->so_buffers
,
4215 4 * 4 * GENX(3DSTATE_SO_BUFFER_length
));
4216 for (int i
= 0; i
< 4; i
++) {
4217 struct iris_stream_output_target
*tgt
=
4218 (void *) ice
->state
.so_target
[i
];
4220 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->base
.buffer
),
4222 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->offset
.res
),
4228 if ((dirty
& IRIS_DIRTY_SO_DECL_LIST
) && ice
->state
.streamout
) {
4229 uint32_t *decl_list
=
4230 ice
->state
.streamout
+ GENX(3DSTATE_STREAMOUT_length
);
4231 iris_batch_emit(batch
, decl_list
, 4 * ((decl_list
[0] & 0xff) + 2));
4234 if (dirty
& IRIS_DIRTY_STREAMOUT
) {
4235 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
4237 uint32_t dynamic_sol
[GENX(3DSTATE_STREAMOUT_length
)];
4238 iris_pack_command(GENX(3DSTATE_STREAMOUT
), dynamic_sol
, sol
) {
4239 sol
.SOFunctionEnable
= true;
4240 sol
.SOStatisticsEnable
= true;
4242 sol
.RenderingDisable
= cso_rast
->rasterizer_discard
&&
4243 !ice
->state
.prims_generated_query_active
;
4244 sol
.ReorderMode
= cso_rast
->flatshade_first
? LEADING
: TRAILING
;
4247 assert(ice
->state
.streamout
);
4249 iris_emit_merge(batch
, ice
->state
.streamout
, dynamic_sol
,
4250 GENX(3DSTATE_STREAMOUT_length
));
4253 if (dirty
& IRIS_DIRTY_STREAMOUT
) {
4254 iris_emit_cmd(batch
, GENX(3DSTATE_STREAMOUT
), sol
);
4258 if (dirty
& IRIS_DIRTY_CLIP
) {
4259 struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
4260 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4262 uint32_t dynamic_clip
[GENX(3DSTATE_CLIP_length
)];
4263 iris_pack_command(GENX(3DSTATE_CLIP
), &dynamic_clip
, cl
) {
4264 cl
.StatisticsEnable
= ice
->state
.statistics_counters_enabled
;
4265 cl
.ClipMode
= cso_rast
->rasterizer_discard
? CLIPMODE_REJECT_ALL
4267 if (wm_prog_data
->barycentric_interp_modes
&
4268 BRW_BARYCENTRIC_NONPERSPECTIVE_BITS
)
4269 cl
.NonPerspectiveBarycentricEnable
= true;
4271 cl
.ForceZeroRTAIndexEnable
= cso_fb
->layers
== 0;
4272 cl
.MaximumVPIndex
= ice
->state
.num_viewports
- 1;
4274 iris_emit_merge(batch
, cso_rast
->clip
, dynamic_clip
,
4275 ARRAY_SIZE(cso_rast
->clip
));
4278 if (dirty
& IRIS_DIRTY_RASTER
) {
4279 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4280 iris_batch_emit(batch
, cso
->raster
, sizeof(cso
->raster
));
4281 iris_batch_emit(batch
, cso
->sf
, sizeof(cso
->sf
));
4285 /* XXX: FS program updates needs to flag IRIS_DIRTY_WM */
4286 if (dirty
& IRIS_DIRTY_WM
) {
4287 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4288 uint32_t dynamic_wm
[GENX(3DSTATE_WM_length
)];
4290 iris_pack_command(GENX(3DSTATE_WM
), &dynamic_wm
, wm
) {
4291 wm
.StatisticsEnable
= ice
->state
.statistics_counters_enabled
;
4293 wm
.BarycentricInterpolationMode
=
4294 wm_prog_data
->barycentric_interp_modes
;
4296 if (wm_prog_data
->early_fragment_tests
)
4297 wm
.EarlyDepthStencilControl
= EDSC_PREPS
;
4298 else if (wm_prog_data
->has_side_effects
)
4299 wm
.EarlyDepthStencilControl
= EDSC_PSEXEC
;
4301 iris_emit_merge(batch
, cso
->wm
, dynamic_wm
, ARRAY_SIZE(cso
->wm
));
4304 if (dirty
& IRIS_DIRTY_SBE
) {
4305 iris_emit_sbe(batch
, ice
);
4308 if (dirty
& IRIS_DIRTY_PS_BLEND
) {
4309 struct iris_blend_state
*cso_blend
= ice
->state
.cso_blend
;
4310 struct iris_depth_stencil_alpha_state
*cso_zsa
= ice
->state
.cso_zsa
;
4311 uint32_t dynamic_pb
[GENX(3DSTATE_PS_BLEND_length
)];
4312 iris_pack_command(GENX(3DSTATE_PS_BLEND
), &dynamic_pb
, pb
) {
4313 pb
.HasWriteableRT
= true; // XXX: comes from somewhere :(
4314 pb
.AlphaTestEnable
= cso_zsa
->alpha
.enabled
;
4317 iris_emit_merge(batch
, cso_blend
->ps_blend
, dynamic_pb
,
4318 ARRAY_SIZE(cso_blend
->ps_blend
));
4321 if (dirty
& IRIS_DIRTY_WM_DEPTH_STENCIL
) {
4322 struct iris_depth_stencil_alpha_state
*cso
= ice
->state
.cso_zsa
;
4323 struct pipe_stencil_ref
*p_stencil_refs
= &ice
->state
.stencil_ref
;
4325 uint32_t stencil_refs
[GENX(3DSTATE_WM_DEPTH_STENCIL_length
)];
4326 iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL
), &stencil_refs
, wmds
) {
4327 wmds
.StencilReferenceValue
= p_stencil_refs
->ref_value
[0];
4328 wmds
.BackfaceStencilReferenceValue
= p_stencil_refs
->ref_value
[1];
4330 iris_emit_merge(batch
, cso
->wmds
, stencil_refs
, ARRAY_SIZE(cso
->wmds
));
4333 if (dirty
& IRIS_DIRTY_SCISSOR_RECT
) {
4334 uint32_t scissor_offset
=
4335 emit_state(batch
, ice
->state
.dynamic_uploader
,
4336 &ice
->state
.last_res
.scissor
,
4337 ice
->state
.scissors
,
4338 sizeof(struct pipe_scissor_state
) *
4339 ice
->state
.num_viewports
, 32);
4341 iris_emit_cmd(batch
, GENX(3DSTATE_SCISSOR_STATE_POINTERS
), ptr
) {
4342 ptr
.ScissorRectPointer
= scissor_offset
;
4346 if (dirty
& IRIS_DIRTY_DEPTH_BUFFER
) {
4347 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4348 struct iris_depth_buffer_state
*cso_z
= &ice
->state
.genx
->depth_buffer
;
4350 iris_batch_emit(batch
, cso_z
->packets
, sizeof(cso_z
->packets
));
4352 if (cso_fb
->zsbuf
) {
4353 struct iris_resource
*zres
, *sres
;
4354 iris_get_depth_stencil_resources(cso_fb
->zsbuf
->texture
,
4356 // XXX: might not be writable...
4358 iris_use_pinned_bo(batch
, zres
->bo
, true);
4360 iris_use_pinned_bo(batch
, sres
->bo
, true);
4364 if (dirty
& IRIS_DIRTY_POLYGON_STIPPLE
) {
4365 iris_emit_cmd(batch
, GENX(3DSTATE_POLY_STIPPLE_PATTERN
), poly
) {
4366 for (int i
= 0; i
< 32; i
++) {
4367 poly
.PatternRow
[i
] = ice
->state
.poly_stipple
.stipple
[i
];
4372 if (dirty
& IRIS_DIRTY_LINE_STIPPLE
) {
4373 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4374 iris_batch_emit(batch
, cso
->line_stipple
, sizeof(cso
->line_stipple
));
4377 if (dirty
& IRIS_DIRTY_VF_TOPOLOGY
) {
4378 iris_emit_cmd(batch
, GENX(3DSTATE_VF_TOPOLOGY
), topo
) {
4379 topo
.PrimitiveTopologyType
=
4380 translate_prim_type(draw
->mode
, draw
->vertices_per_patch
);
4384 if (dirty
& IRIS_DIRTY_VERTEX_BUFFERS
) {
4385 int count
= util_bitcount64(ice
->state
.bound_vertex_buffers
);
4388 /* The VF cache designers cut corners, and made the cache key's
4389 * <VertexBufferIndex, Memory Address> tuple only consider the bottom
4390 * 32 bits of the address. If you have two vertex buffers which get
4391 * placed exactly 4 GiB apart and use them in back-to-back draw calls,
4392 * you can get collisions (even within a single batch).
4394 * So, we need to do a VF cache invalidate if the buffer for a VB
4395 * slot slot changes [48:32] address bits from the previous time.
4397 unsigned flush_flags
= 0;
4399 uint64_t bound
= ice
->state
.bound_vertex_buffers
;
4401 const int i
= u_bit_scan64(&bound
);
4402 uint16_t high_bits
= 0;
4404 struct iris_resource
*res
=
4405 (void *) genx
->vertex_buffers
[i
].resource
;
4407 iris_use_pinned_bo(batch
, res
->bo
, false);
4409 high_bits
= res
->bo
->gtt_offset
>> 32ull;
4410 if (high_bits
!= ice
->state
.last_vbo_high_bits
[i
]) {
4411 flush_flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
4412 ice
->state
.last_vbo_high_bits
[i
] = high_bits
;
4415 /* If the buffer was written to by streamout, we may need
4416 * to stall so those writes land and become visible to the
4419 * TODO: This may stall more than necessary.
4421 if (res
->bind_history
& PIPE_BIND_STREAM_OUTPUT
)
4422 flush_flags
|= PIPE_CONTROL_CS_STALL
;
4427 iris_emit_pipe_control_flush(batch
, flush_flags
);
4429 const unsigned vb_dwords
= GENX(VERTEX_BUFFER_STATE_length
);
4432 iris_get_command_space(batch
, 4 * (1 + vb_dwords
* count
));
4433 _iris_pack_command(batch
, GENX(3DSTATE_VERTEX_BUFFERS
), map
, vb
) {
4434 vb
.DWordLength
= (vb_dwords
* count
+ 1) - 2;
4438 bound
= ice
->state
.bound_vertex_buffers
;
4440 const int i
= u_bit_scan64(&bound
);
4441 memcpy(map
, genx
->vertex_buffers
[i
].state
,
4442 sizeof(uint32_t) * vb_dwords
);
4448 if (dirty
& IRIS_DIRTY_VERTEX_ELEMENTS
) {
4449 struct iris_vertex_element_state
*cso
= ice
->state
.cso_vertex_elements
;
4450 const unsigned entries
= MAX2(cso
->count
, 1);
4451 iris_batch_emit(batch
, cso
->vertex_elements
, sizeof(uint32_t) *
4452 (1 + entries
* GENX(VERTEX_ELEMENT_STATE_length
)));
4453 iris_batch_emit(batch
, cso
->vf_instancing
, sizeof(uint32_t) *
4454 entries
* GENX(3DSTATE_VF_INSTANCING_length
));
4457 if (dirty
& IRIS_DIRTY_VF_SGVS
) {
4458 const struct brw_vs_prog_data
*vs_prog_data
= (void *)
4459 ice
->shaders
.prog
[MESA_SHADER_VERTEX
]->prog_data
;
4460 struct iris_vertex_element_state
*cso
= ice
->state
.cso_vertex_elements
;
4462 iris_emit_cmd(batch
, GENX(3DSTATE_VF_SGVS
), sgv
) {
4463 if (vs_prog_data
->uses_vertexid
) {
4464 sgv
.VertexIDEnable
= true;
4465 sgv
.VertexIDComponentNumber
= 2;
4466 sgv
.VertexIDElementOffset
= cso
->count
;
4469 if (vs_prog_data
->uses_instanceid
) {
4470 sgv
.InstanceIDEnable
= true;
4471 sgv
.InstanceIDComponentNumber
= 3;
4472 sgv
.InstanceIDElementOffset
= cso
->count
;
4477 if (dirty
& IRIS_DIRTY_VF
) {
4478 iris_emit_cmd(batch
, GENX(3DSTATE_VF
), vf
) {
4479 if (draw
->primitive_restart
) {
4480 vf
.IndexedDrawCutIndexEnable
= true;
4481 vf
.CutIndex
= draw
->restart_index
;
4486 // XXX: Gen8 - PMA fix
4490 iris_upload_render_state(struct iris_context
*ice
,
4491 struct iris_batch
*batch
,
4492 const struct pipe_draw_info
*draw
)
4494 /* Always pin the binder. If we're emitting new binding table pointers,
4495 * we need it. If not, we're probably inheriting old tables via the
4496 * context, and need it anyway. Since true zero-bindings cases are
4497 * practically non-existent, just pin it and avoid last_res tracking.
4499 iris_use_pinned_bo(batch
, ice
->state
.binder
.bo
, false);
4501 if (!batch
->contains_draw
) {
4502 iris_restore_render_saved_bos(ice
, batch
, draw
);
4503 batch
->contains_draw
= true;
4506 iris_upload_dirty_render_state(ice
, batch
, draw
);
4508 if (draw
->index_size
> 0) {
4511 if (draw
->has_user_indices
) {
4512 u_upload_data(ice
->ctx
.stream_uploader
, 0,
4513 draw
->count
* draw
->index_size
, 4, draw
->index
.user
,
4514 &offset
, &ice
->state
.last_res
.index_buffer
);
4516 struct iris_resource
*res
= (void *) draw
->index
.resource
;
4517 res
->bind_history
|= PIPE_BIND_INDEX_BUFFER
;
4519 pipe_resource_reference(&ice
->state
.last_res
.index_buffer
,
4520 draw
->index
.resource
);
4524 struct iris_bo
*bo
= iris_resource_bo(ice
->state
.last_res
.index_buffer
);
4526 iris_emit_cmd(batch
, GENX(3DSTATE_INDEX_BUFFER
), ib
) {
4527 ib
.IndexFormat
= draw
->index_size
>> 1;
4529 ib
.BufferSize
= bo
->size
;
4530 ib
.BufferStartingAddress
= ro_bo(bo
, offset
);
4533 /* The VF cache key only uses 32-bits, see vertex buffer comment above */
4534 uint16_t high_bits
= bo
->gtt_offset
>> 32ull;
4535 if (high_bits
!= ice
->state
.last_index_bo_high_bits
) {
4536 iris_emit_pipe_control_flush(batch
, PIPE_CONTROL_VF_CACHE_INVALIDATE
);
4537 ice
->state
.last_index_bo_high_bits
= high_bits
;
4541 #define _3DPRIM_END_OFFSET 0x2420
4542 #define _3DPRIM_START_VERTEX 0x2430
4543 #define _3DPRIM_VERTEX_COUNT 0x2434
4544 #define _3DPRIM_INSTANCE_COUNT 0x2438
4545 #define _3DPRIM_START_INSTANCE 0x243C
4546 #define _3DPRIM_BASE_VERTEX 0x2440
4548 if (draw
->indirect
) {
4549 /* We don't support this MultidrawIndirect. */
4550 assert(!draw
->indirect
->indirect_draw_count
);
4552 struct iris_bo
*bo
= iris_resource_bo(draw
->indirect
->buffer
);
4555 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4556 lrm
.RegisterAddress
= _3DPRIM_VERTEX_COUNT
;
4557 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 0);
4559 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4560 lrm
.RegisterAddress
= _3DPRIM_INSTANCE_COUNT
;
4561 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 4);
4563 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4564 lrm
.RegisterAddress
= _3DPRIM_START_VERTEX
;
4565 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 8);
4567 if (draw
->index_size
) {
4568 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4569 lrm
.RegisterAddress
= _3DPRIM_BASE_VERTEX
;
4570 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 12);
4572 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4573 lrm
.RegisterAddress
= _3DPRIM_START_INSTANCE
;
4574 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 16);
4577 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4578 lrm
.RegisterAddress
= _3DPRIM_START_INSTANCE
;
4579 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 12);
4581 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
4582 lri
.RegisterOffset
= _3DPRIM_BASE_VERTEX
;
4588 iris_emit_cmd(batch
, GENX(3DPRIMITIVE
), prim
) {
4589 prim
.StartInstanceLocation
= draw
->start_instance
;
4590 prim
.InstanceCount
= draw
->instance_count
;
4591 prim
.VertexCountPerInstance
= draw
->count
;
4592 prim
.VertexAccessType
= draw
->index_size
> 0 ? RANDOM
: SEQUENTIAL
;
4593 prim
.PredicateEnable
=
4594 ice
->state
.predicate
== IRIS_PREDICATE_STATE_USE_BIT
;
4596 // XXX: this is probably bonkers.
4597 prim
.StartVertexLocation
= draw
->start
;
4599 prim
.IndirectParameterEnable
= draw
->indirect
!= NULL
;
4601 if (draw
->index_size
) {
4602 prim
.BaseVertexLocation
+= draw
->index_bias
;
4604 prim
.StartVertexLocation
+= draw
->index_bias
;
4607 //prim.BaseVertexLocation = ...;
4612 iris_upload_compute_state(struct iris_context
*ice
,
4613 struct iris_batch
*batch
,
4614 const struct pipe_grid_info
*grid
)
4616 const uint64_t dirty
= ice
->state
.dirty
;
4617 struct iris_screen
*screen
= batch
->screen
;
4618 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
4619 struct iris_binder
*binder
= &ice
->state
.binder
;
4620 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_COMPUTE
];
4621 struct iris_compiled_shader
*shader
=
4622 ice
->shaders
.prog
[MESA_SHADER_COMPUTE
];
4623 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
4624 struct brw_cs_prog_data
*cs_prog_data
= (void *) prog_data
;
4626 if ((dirty
& IRIS_DIRTY_CONSTANTS_CS
) && shs
->cbuf0_needs_upload
)
4627 upload_uniforms(ice
, MESA_SHADER_COMPUTE
);
4629 if (dirty
& IRIS_DIRTY_BINDINGS_CS
)
4630 iris_populate_binding_table(ice
, batch
, MESA_SHADER_COMPUTE
, false);
4632 iris_use_optional_res(batch
, shs
->sampler_table
.res
, false);
4633 iris_use_pinned_bo(batch
, iris_resource_bo(shader
->assembly
.res
), false);
4635 if (ice
->state
.need_border_colors
)
4636 iris_use_pinned_bo(batch
, ice
->state
.border_color_pool
.bo
, false);
4638 if (dirty
& IRIS_DIRTY_CS
) {
4639 /* The MEDIA_VFE_STATE documentation for Gen8+ says:
4641 * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
4642 * the only bits that are changed are scoreboard related: Scoreboard
4643 * Enable, Scoreboard Type, Scoreboard Mask, Scoreboard Delta. For
4644 * these scoreboard related states, a MEDIA_STATE_FLUSH is
4647 iris_emit_pipe_control_flush(batch
, PIPE_CONTROL_CS_STALL
);
4649 iris_emit_cmd(batch
, GENX(MEDIA_VFE_STATE
), vfe
) {
4650 if (prog_data
->total_scratch
) {
4651 uint32_t scratch_addr
=
4652 iris_get_scratch_space(ice
, prog_data
->total_scratch
,
4653 MESA_SHADER_COMPUTE
);
4654 vfe
.PerThreadScratchSpace
= ffs(prog_data
->total_scratch
) - 11;
4655 vfe
.ScratchSpaceBasePointer
= rw_bo(NULL
, scratch_addr
);
4658 vfe
.MaximumNumberofThreads
=
4659 devinfo
->max_cs_threads
* screen
->subslice_total
- 1;
4661 vfe
.ResetGatewayTimer
=
4662 Resettingrelativetimerandlatchingtheglobaltimestamp
;
4665 vfe
.NumberofURBEntries
= 2;
4666 vfe
.URBEntryAllocationSize
= 2;
4668 // XXX: Use Indirect Payload Storage?
4669 vfe
.CURBEAllocationSize
=
4670 ALIGN(cs_prog_data
->push
.per_thread
.regs
* cs_prog_data
->threads
+
4671 cs_prog_data
->push
.cross_thread
.regs
, 2);
4675 // XXX: hack iris_set_constant_buffers to upload these thread counts
4676 // XXX: along with regular uniforms for compute shaders, somehow.
4678 uint32_t curbe_data_offset
= 0;
4679 // TODO: Move subgroup-id into uniforms ubo so we can push uniforms
4680 assert(cs_prog_data
->push
.cross_thread
.dwords
== 0 &&
4681 cs_prog_data
->push
.per_thread
.dwords
== 1 &&
4682 cs_prog_data
->base
.param
[0] == BRW_PARAM_BUILTIN_SUBGROUP_ID
);
4683 struct pipe_resource
*curbe_data_res
= NULL
;
4684 uint32_t *curbe_data_map
=
4685 stream_state(batch
, ice
->state
.dynamic_uploader
, &curbe_data_res
,
4686 ALIGN(cs_prog_data
->push
.total
.size
, 64), 64,
4687 &curbe_data_offset
);
4688 assert(curbe_data_map
);
4689 memset(curbe_data_map
, 0x5a, ALIGN(cs_prog_data
->push
.total
.size
, 64));
4690 iris_fill_cs_push_const_buffer(cs_prog_data
, curbe_data_map
);
4692 if (dirty
& IRIS_DIRTY_CONSTANTS_CS
) {
4693 iris_emit_cmd(batch
, GENX(MEDIA_CURBE_LOAD
), curbe
) {
4694 curbe
.CURBETotalDataLength
=
4695 ALIGN(cs_prog_data
->push
.total
.size
, 64);
4696 curbe
.CURBEDataStartAddress
= curbe_data_offset
;
4700 if (dirty
& (IRIS_DIRTY_SAMPLER_STATES_CS
|
4701 IRIS_DIRTY_BINDINGS_CS
|
4702 IRIS_DIRTY_CONSTANTS_CS
|
4704 struct pipe_resource
*desc_res
= NULL
;
4705 uint32_t desc
[GENX(INTERFACE_DESCRIPTOR_DATA_length
)];
4707 iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA
), desc
, idd
) {
4708 idd
.SamplerStatePointer
= shs
->sampler_table
.offset
;
4709 idd
.BindingTablePointer
= binder
->bt_offset
[MESA_SHADER_COMPUTE
];
4712 for (int i
= 0; i
< GENX(INTERFACE_DESCRIPTOR_DATA_length
); i
++)
4713 desc
[i
] |= ((uint32_t *) shader
->derived_data
)[i
];
4715 iris_emit_cmd(batch
, GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD
), load
) {
4716 load
.InterfaceDescriptorTotalLength
=
4717 GENX(INTERFACE_DESCRIPTOR_DATA_length
) * sizeof(uint32_t);
4718 load
.InterfaceDescriptorDataStartAddress
=
4719 emit_state(batch
, ice
->state
.dynamic_uploader
,
4720 &desc_res
, desc
, sizeof(desc
), 32);
4723 pipe_resource_reference(&desc_res
, NULL
);
4726 uint32_t group_size
= grid
->block
[0] * grid
->block
[1] * grid
->block
[2];
4727 uint32_t remainder
= group_size
& (cs_prog_data
->simd_size
- 1);
4728 uint32_t right_mask
;
4731 right_mask
= ~0u >> (32 - remainder
);
4733 right_mask
= ~0u >> (32 - cs_prog_data
->simd_size
);
4735 #define GPGPU_DISPATCHDIMX 0x2500
4736 #define GPGPU_DISPATCHDIMY 0x2504
4737 #define GPGPU_DISPATCHDIMZ 0x2508
4739 if (grid
->indirect
) {
4740 struct iris_state_ref
*grid_size
= &ice
->state
.grid_size
;
4741 struct iris_bo
*bo
= iris_resource_bo(grid_size
->res
);
4742 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4743 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMX
;
4744 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 0);
4746 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4747 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMY
;
4748 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 4);
4750 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4751 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMZ
;
4752 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 8);
4756 iris_emit_cmd(batch
, GENX(GPGPU_WALKER
), ggw
) {
4757 ggw
.IndirectParameterEnable
= grid
->indirect
!= NULL
;
4758 ggw
.SIMDSize
= cs_prog_data
->simd_size
/ 16;
4759 ggw
.ThreadDepthCounterMaximum
= 0;
4760 ggw
.ThreadHeightCounterMaximum
= 0;
4761 ggw
.ThreadWidthCounterMaximum
= cs_prog_data
->threads
- 1;
4762 ggw
.ThreadGroupIDXDimension
= grid
->grid
[0];
4763 ggw
.ThreadGroupIDYDimension
= grid
->grid
[1];
4764 ggw
.ThreadGroupIDZDimension
= grid
->grid
[2];
4765 ggw
.RightExecutionMask
= right_mask
;
4766 ggw
.BottomExecutionMask
= 0xffffffff;
4769 iris_emit_cmd(batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
4771 if (!batch
->contains_draw
) {
4772 iris_restore_compute_saved_bos(ice
, batch
, grid
);
4773 batch
->contains_draw
= true;
4778 * State module teardown.
4781 iris_destroy_state(struct iris_context
*ice
)
4783 struct iris_genx_state
*genx
= ice
->state
.genx
;
4785 uint64_t bound_vbs
= ice
->state
.bound_vertex_buffers
;
4787 const int i
= u_bit_scan64(&bound_vbs
);
4788 pipe_resource_reference(&genx
->vertex_buffers
[i
].resource
, NULL
);
4791 // XXX: unreference resources/surfaces.
4792 for (unsigned i
= 0; i
< ice
->state
.framebuffer
.nr_cbufs
; i
++) {
4793 pipe_surface_reference(&ice
->state
.framebuffer
.cbufs
[i
], NULL
);
4795 pipe_surface_reference(&ice
->state
.framebuffer
.zsbuf
, NULL
);
4797 for (int stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
4798 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4799 pipe_resource_reference(&shs
->sampler_table
.res
, NULL
);
4801 free(ice
->state
.genx
);
4803 pipe_resource_reference(&ice
->state
.unbound_tex
.res
, NULL
);
4805 pipe_resource_reference(&ice
->state
.last_res
.cc_vp
, NULL
);
4806 pipe_resource_reference(&ice
->state
.last_res
.sf_cl_vp
, NULL
);
4807 pipe_resource_reference(&ice
->state
.last_res
.color_calc
, NULL
);
4808 pipe_resource_reference(&ice
->state
.last_res
.scissor
, NULL
);
4809 pipe_resource_reference(&ice
->state
.last_res
.blend
, NULL
);
4810 pipe_resource_reference(&ice
->state
.last_res
.index_buffer
, NULL
);
4813 /* ------------------------------------------------------------------- */
4816 iris_load_register_reg32(struct iris_batch
*batch
, uint32_t dst
,
4819 _iris_emit_lrr(batch
, dst
, src
);
4823 iris_load_register_reg64(struct iris_batch
*batch
, uint32_t dst
,
4826 _iris_emit_lrr(batch
, dst
, src
);
4827 _iris_emit_lrr(batch
, dst
+ 4, src
+ 4);
4831 iris_load_register_imm32(struct iris_batch
*batch
, uint32_t reg
,
4834 _iris_emit_lri(batch
, reg
, val
);
4838 iris_load_register_imm64(struct iris_batch
*batch
, uint32_t reg
,
4841 _iris_emit_lri(batch
, reg
+ 0, val
& 0xffffffff);
4842 _iris_emit_lri(batch
, reg
+ 4, val
>> 32);
4846 * Emit MI_LOAD_REGISTER_MEM to load a 32-bit MMIO register from a buffer.
4849 iris_load_register_mem32(struct iris_batch
*batch
, uint32_t reg
,
4850 struct iris_bo
*bo
, uint32_t offset
)
4852 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4853 lrm
.RegisterAddress
= reg
;
4854 lrm
.MemoryAddress
= ro_bo(bo
, offset
);
4859 * Load a 64-bit value from a buffer into a MMIO register via
4860 * two MI_LOAD_REGISTER_MEM commands.
4863 iris_load_register_mem64(struct iris_batch
*batch
, uint32_t reg
,
4864 struct iris_bo
*bo
, uint32_t offset
)
4866 iris_load_register_mem32(batch
, reg
+ 0, bo
, offset
+ 0);
4867 iris_load_register_mem32(batch
, reg
+ 4, bo
, offset
+ 4);
4871 iris_store_register_mem32(struct iris_batch
*batch
, uint32_t reg
,
4872 struct iris_bo
*bo
, uint32_t offset
,
4875 iris_emit_cmd(batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
4876 srm
.RegisterAddress
= reg
;
4877 srm
.MemoryAddress
= rw_bo(bo
, offset
);
4878 srm
.PredicateEnable
= predicated
;
4883 iris_store_register_mem64(struct iris_batch
*batch
, uint32_t reg
,
4884 struct iris_bo
*bo
, uint32_t offset
,
4887 iris_store_register_mem32(batch
, reg
+ 0, bo
, offset
+ 0, predicated
);
4888 iris_store_register_mem32(batch
, reg
+ 4, bo
, offset
+ 4, predicated
);
4892 iris_store_data_imm32(struct iris_batch
*batch
,
4893 struct iris_bo
*bo
, uint32_t offset
,
4896 iris_emit_cmd(batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
4897 sdi
.Address
= rw_bo(bo
, offset
);
4898 sdi
.ImmediateData
= imm
;
4903 iris_store_data_imm64(struct iris_batch
*batch
,
4904 struct iris_bo
*bo
, uint32_t offset
,
4907 /* Can't use iris_emit_cmd because MI_STORE_DATA_IMM has a length of
4908 * 2 in genxml but it's actually variable length and we need 5 DWords.
4910 void *map
= iris_get_command_space(batch
, 4 * 5);
4911 _iris_pack_command(batch
, GENX(MI_STORE_DATA_IMM
), map
, sdi
) {
4912 sdi
.DWordLength
= 5 - 2;
4913 sdi
.Address
= rw_bo(bo
, offset
);
4914 sdi
.ImmediateData
= imm
;
4919 iris_copy_mem_mem(struct iris_batch
*batch
,
4920 struct iris_bo
*dst_bo
, uint32_t dst_offset
,
4921 struct iris_bo
*src_bo
, uint32_t src_offset
,
4924 /* MI_COPY_MEM_MEM operates on DWords. */
4925 assert(bytes
% 4 == 0);
4926 assert(dst_offset
% 4 == 0);
4927 assert(src_offset
% 4 == 0);
4929 for (unsigned i
= 0; i
< bytes
; i
+= 4) {
4930 iris_emit_cmd(batch
, GENX(MI_COPY_MEM_MEM
), cp
) {
4931 cp
.DestinationMemoryAddress
= rw_bo(dst_bo
, dst_offset
+ i
);
4932 cp
.SourceMemoryAddress
= ro_bo(src_bo
, src_offset
+ i
);
4937 /* ------------------------------------------------------------------- */
4940 flags_to_post_sync_op(uint32_t flags
)
4942 if (flags
& PIPE_CONTROL_WRITE_IMMEDIATE
)
4943 return WriteImmediateData
;
4945 if (flags
& PIPE_CONTROL_WRITE_DEPTH_COUNT
)
4946 return WritePSDepthCount
;
4948 if (flags
& PIPE_CONTROL_WRITE_TIMESTAMP
)
4949 return WriteTimestamp
;
4955 * Do the given flags have a Post Sync or LRI Post Sync operation?
4957 static enum pipe_control_flags
4958 get_post_sync_flags(enum pipe_control_flags flags
)
4960 flags
&= PIPE_CONTROL_WRITE_IMMEDIATE
|
4961 PIPE_CONTROL_WRITE_DEPTH_COUNT
|
4962 PIPE_CONTROL_WRITE_TIMESTAMP
|
4963 PIPE_CONTROL_LRI_POST_SYNC_OP
;
4965 /* Only one "Post Sync Op" is allowed, and it's mutually exclusive with
4966 * "LRI Post Sync Operation". So more than one bit set would be illegal.
4968 assert(util_bitcount(flags
) <= 1);
4973 #define IS_COMPUTE_PIPELINE(batch) (batch->name == IRIS_BATCH_COMPUTE)
4976 * Emit a series of PIPE_CONTROL commands, taking into account any
4977 * workarounds necessary to actually accomplish the caller's request.
4979 * Unless otherwise noted, spec quotations in this function come from:
4981 * Synchronization of the 3D Pipeline > PIPE_CONTROL Command > Programming
4982 * Restrictions for PIPE_CONTROL.
4984 * You should not use this function directly. Use the helpers in
4985 * iris_pipe_control.c instead, which may split the pipe control further.
4988 iris_emit_raw_pipe_control(struct iris_batch
*batch
, uint32_t flags
,
4989 struct iris_bo
*bo
, uint32_t offset
, uint64_t imm
)
4991 UNUSED
const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
4992 enum pipe_control_flags post_sync_flags
= get_post_sync_flags(flags
);
4993 enum pipe_control_flags non_lri_post_sync_flags
=
4994 post_sync_flags
& ~PIPE_CONTROL_LRI_POST_SYNC_OP
;
4996 /* Recursive PIPE_CONTROL workarounds --------------------------------
4997 * (http://knowyourmeme.com/memes/xzibit-yo-dawg)
4999 * We do these first because we want to look at the original operation,
5000 * rather than any workarounds we set.
5002 if (GEN_GEN
== 9 && (flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
)) {
5003 /* The PIPE_CONTROL "VF Cache Invalidation Enable" bit description
5004 * lists several workarounds:
5006 * "Project: SKL, KBL, BXT
5008 * If the VF Cache Invalidation Enable is set to a 1 in a
5009 * PIPE_CONTROL, a separate Null PIPE_CONTROL, all bitfields
5010 * sets to 0, with the VF Cache Invalidation Enable set to 0
5011 * needs to be sent prior to the PIPE_CONTROL with VF Cache
5012 * Invalidation Enable set to a 1."
5014 iris_emit_raw_pipe_control(batch
, 0, NULL
, 0, 0);
5017 if (GEN_GEN
== 9 && IS_COMPUTE_PIPELINE(batch
) && post_sync_flags
) {
5018 /* Project: SKL / Argument: LRI Post Sync Operation [23]
5020 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
5021 * programmed prior to programming a PIPECONTROL command with "LRI
5022 * Post Sync Operation" in GPGPU mode of operation (i.e when
5023 * PIPELINE_SELECT command is set to GPGPU mode of operation)."
5025 * The same text exists a few rows below for Post Sync Op.
5027 iris_emit_raw_pipe_control(batch
, PIPE_CONTROL_CS_STALL
, bo
, offset
, imm
);
5030 if (GEN_GEN
== 10 && (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
)) {
5032 * "Before sending a PIPE_CONTROL command with bit 12 set, SW must issue
5033 * another PIPE_CONTROL with Render Target Cache Flush Enable (bit 12)
5034 * = 0 and Pipe Control Flush Enable (bit 7) = 1"
5036 iris_emit_raw_pipe_control(batch
, PIPE_CONTROL_FLUSH_ENABLE
, bo
,
5040 /* "Flush Types" workarounds ---------------------------------------------
5041 * We do these now because they may add post-sync operations or CS stalls.
5044 if (GEN_GEN
< 11 && flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
) {
5045 /* Project: BDW, SKL+ (stopping at CNL) / Argument: VF Invalidate
5047 * "'Post Sync Operation' must be enabled to 'Write Immediate Data' or
5048 * 'Write PS Depth Count' or 'Write Timestamp'."
5051 flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
5052 post_sync_flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
5053 non_lri_post_sync_flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
5054 bo
= batch
->screen
->workaround_bo
;
5058 /* #1130 from Gen10 workarounds page:
5060 * "Enable Depth Stall on every Post Sync Op if Render target Cache
5061 * Flush is not enabled in same PIPE CONTROL and Enable Pixel score
5062 * board stall if Render target cache flush is enabled."
5064 * Applicable to CNL B0 and C0 steppings only.
5066 * The wording here is unclear, and this workaround doesn't look anything
5067 * like the internal bug report recommendations, but leave it be for now...
5069 if (GEN_GEN
== 10) {
5070 if (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
) {
5071 flags
|= PIPE_CONTROL_STALL_AT_SCOREBOARD
;
5072 } else if (flags
& non_lri_post_sync_flags
) {
5073 flags
|= PIPE_CONTROL_DEPTH_STALL
;
5077 if (flags
& PIPE_CONTROL_DEPTH_STALL
) {
5078 /* From the PIPE_CONTROL instruction table, bit 13 (Depth Stall Enable):
5080 * "This bit must be DISABLED for operations other than writing
5083 * This seems like nonsense. An Ivybridge workaround requires us to
5084 * emit a PIPE_CONTROL with a depth stall and write immediate post-sync
5085 * operation. Gen8+ requires us to emit depth stalls and depth cache
5086 * flushes together. So, it's hard to imagine this means anything other
5087 * than "we originally intended this to be used for PS_DEPTH_COUNT".
5089 * We ignore the supposed restriction and do nothing.
5093 if (flags
& (PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5094 PIPE_CONTROL_STALL_AT_SCOREBOARD
)) {
5095 /* From the PIPE_CONTROL instruction table, bit 12 and bit 1:
5097 * "This bit must be DISABLED for End-of-pipe (Read) fences,
5098 * PS_DEPTH_COUNT or TIMESTAMP queries."
5100 * TODO: Implement end-of-pipe checking.
5102 assert(!(post_sync_flags
& (PIPE_CONTROL_WRITE_DEPTH_COUNT
|
5103 PIPE_CONTROL_WRITE_TIMESTAMP
)));
5106 if (GEN_GEN
< 11 && (flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
)) {
5107 /* From the PIPE_CONTROL instruction table, bit 1:
5109 * "This bit is ignored if Depth Stall Enable is set.
5110 * Further, the render cache is not flushed even if Write Cache
5111 * Flush Enable bit is set."
5113 * We assert that the caller doesn't do this combination, to try and
5114 * prevent mistakes. It shouldn't hurt the GPU, though.
5116 * We skip this check on Gen11+ as the "Stall at Pixel Scoreboard"
5117 * and "Render Target Flush" combo is explicitly required for BTI
5118 * update workarounds.
5120 assert(!(flags
& (PIPE_CONTROL_DEPTH_STALL
|
5121 PIPE_CONTROL_RENDER_TARGET_FLUSH
)));
5124 /* PIPE_CONTROL page workarounds ------------------------------------- */
5126 if (GEN_GEN
<= 8 && (flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
)) {
5127 /* From the PIPE_CONTROL page itself:
5130 * Restriction: Pipe_control with CS-stall bit set must be issued
5131 * before a pipe-control command that has the State Cache
5132 * Invalidate bit set."
5134 flags
|= PIPE_CONTROL_CS_STALL
;
5137 if (flags
& PIPE_CONTROL_FLUSH_LLC
) {
5138 /* From the PIPE_CONTROL instruction table, bit 26 (Flush LLC):
5141 * SW must always program Post-Sync Operation to "Write Immediate
5142 * Data" when Flush LLC is set."
5144 * For now, we just require the caller to do it.
5146 assert(flags
& PIPE_CONTROL_WRITE_IMMEDIATE
);
5149 /* "Post-Sync Operation" workarounds -------------------------------- */
5151 /* Project: All / Argument: Global Snapshot Count Reset [19]
5153 * "This bit must not be exercised on any product.
5154 * Requires stall bit ([20] of DW1) set."
5156 * We don't use this, so we just assert that it isn't used. The
5157 * PIPE_CONTROL instruction page indicates that they intended this
5158 * as a debug feature and don't think it is useful in production,
5159 * but it may actually be usable, should we ever want to.
5161 assert((flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
) == 0);
5163 if (flags
& (PIPE_CONTROL_MEDIA_STATE_CLEAR
|
5164 PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
)) {
5165 /* Project: All / Arguments:
5167 * - Generic Media State Clear [16]
5168 * - Indirect State Pointers Disable [16]
5170 * "Requires stall bit ([20] of DW1) set."
5172 * Also, the PIPE_CONTROL instruction table, bit 16 (Generic Media
5173 * State Clear) says:
5175 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
5176 * programmed prior to programming a PIPECONTROL command with "Media
5177 * State Clear" set in GPGPU mode of operation"
5179 * This is a subset of the earlier rule, so there's nothing to do.
5181 flags
|= PIPE_CONTROL_CS_STALL
;
5184 if (flags
& PIPE_CONTROL_STORE_DATA_INDEX
) {
5185 /* Project: All / Argument: Store Data Index
5187 * "Post-Sync Operation ([15:14] of DW1) must be set to something other
5190 * For now, we just assert that the caller does this. We might want to
5191 * automatically add a write to the workaround BO...
5193 assert(non_lri_post_sync_flags
!= 0);
5196 if (flags
& PIPE_CONTROL_SYNC_GFDT
) {
5197 /* Project: All / Argument: Sync GFDT
5199 * "Post-Sync Operation ([15:14] of DW1) must be set to something other
5200 * than '0' or 0x2520[13] must be set."
5202 * For now, we just assert that the caller does this.
5204 assert(non_lri_post_sync_flags
!= 0);
5207 if (flags
& PIPE_CONTROL_TLB_INVALIDATE
) {
5208 /* Project: IVB+ / Argument: TLB inv
5210 * "Requires stall bit ([20] of DW1) set."
5212 * Also, from the PIPE_CONTROL instruction table:
5215 * Post Sync Operation or CS stall must be set to ensure a TLB
5216 * invalidation occurs. Otherwise no cycle will occur to the TLB
5217 * cache to invalidate."
5219 * This is not a subset of the earlier rule, so there's nothing to do.
5221 flags
|= PIPE_CONTROL_CS_STALL
;
5224 if (GEN_GEN
== 9 && devinfo
->gt
== 4) {
5225 /* TODO: The big Skylake GT4 post sync op workaround */
5228 /* "GPGPU specific workarounds" (both post-sync and flush) ------------ */
5230 if (IS_COMPUTE_PIPELINE(batch
)) {
5231 if (GEN_GEN
>= 9 && (flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
)) {
5232 /* Project: SKL+ / Argument: Tex Invalidate
5233 * "Requires stall bit ([20] of DW) set for all GPGPU Workloads."
5235 flags
|= PIPE_CONTROL_CS_STALL
;
5238 if (GEN_GEN
== 8 && (post_sync_flags
||
5239 (flags
& (PIPE_CONTROL_NOTIFY_ENABLE
|
5240 PIPE_CONTROL_DEPTH_STALL
|
5241 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5242 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
5243 PIPE_CONTROL_DATA_CACHE_FLUSH
)))) {
5244 /* Project: BDW / Arguments:
5246 * - LRI Post Sync Operation [23]
5247 * - Post Sync Op [15:14]
5249 * - Depth Stall [13]
5250 * - Render Target Cache Flush [12]
5251 * - Depth Cache Flush [0]
5252 * - DC Flush Enable [5]
5254 * "Requires stall bit ([20] of DW) set for all GPGPU and Media
5257 flags
|= PIPE_CONTROL_CS_STALL
;
5259 /* Also, from the PIPE_CONTROL instruction table, bit 20:
5262 * This bit must be always set when PIPE_CONTROL command is
5263 * programmed by GPGPU and MEDIA workloads, except for the cases
5264 * when only Read Only Cache Invalidation bits are set (State
5265 * Cache Invalidation Enable, Instruction cache Invalidation
5266 * Enable, Texture Cache Invalidation Enable, Constant Cache
5267 * Invalidation Enable). This is to WA FFDOP CG issue, this WA
5268 * need not implemented when FF_DOP_CG is disable via "Fixed
5269 * Function DOP Clock Gate Disable" bit in RC_PSMI_CTRL register."
5271 * It sounds like we could avoid CS stalls in some cases, but we
5272 * don't currently bother. This list isn't exactly the list above,
5278 /* "Stall" workarounds ----------------------------------------------
5279 * These have to come after the earlier ones because we may have added
5280 * some additional CS stalls above.
5283 if (GEN_GEN
< 9 && (flags
& PIPE_CONTROL_CS_STALL
)) {
5284 /* Project: PRE-SKL, VLV, CHV
5286 * "[All Stepping][All SKUs]:
5288 * One of the following must also be set:
5290 * - Render Target Cache Flush Enable ([12] of DW1)
5291 * - Depth Cache Flush Enable ([0] of DW1)
5292 * - Stall at Pixel Scoreboard ([1] of DW1)
5293 * - Depth Stall ([13] of DW1)
5294 * - Post-Sync Operation ([13] of DW1)
5295 * - DC Flush Enable ([5] of DW1)"
5297 * If we don't already have one of those bits set, we choose to add
5298 * "Stall at Pixel Scoreboard". Some of the other bits require a
5299 * CS stall as a workaround (see above), which would send us into
5300 * an infinite recursion of PIPE_CONTROLs. "Stall at Pixel Scoreboard"
5301 * appears to be safe, so we choose that.
5303 const uint32_t wa_bits
= PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5304 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
5305 PIPE_CONTROL_WRITE_IMMEDIATE
|
5306 PIPE_CONTROL_WRITE_DEPTH_COUNT
|
5307 PIPE_CONTROL_WRITE_TIMESTAMP
|
5308 PIPE_CONTROL_STALL_AT_SCOREBOARD
|
5309 PIPE_CONTROL_DEPTH_STALL
|
5310 PIPE_CONTROL_DATA_CACHE_FLUSH
;
5311 if (!(flags
& wa_bits
))
5312 flags
|= PIPE_CONTROL_STALL_AT_SCOREBOARD
;
5315 /* Emit --------------------------------------------------------------- */
5317 iris_emit_cmd(batch
, GENX(PIPE_CONTROL
), pc
) {
5318 pc
.LRIPostSyncOperation
= NoLRIOperation
;
5319 pc
.PipeControlFlushEnable
= flags
& PIPE_CONTROL_FLUSH_ENABLE
;
5320 pc
.DCFlushEnable
= flags
& PIPE_CONTROL_DATA_CACHE_FLUSH
;
5321 pc
.StoreDataIndex
= 0;
5322 pc
.CommandStreamerStallEnable
= flags
& PIPE_CONTROL_CS_STALL
;
5323 pc
.GlobalSnapshotCountReset
=
5324 flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
;
5325 pc
.TLBInvalidate
= flags
& PIPE_CONTROL_TLB_INVALIDATE
;
5326 pc
.GenericMediaStateClear
= flags
& PIPE_CONTROL_MEDIA_STATE_CLEAR
;
5327 pc
.StallAtPixelScoreboard
= flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
;
5328 pc
.RenderTargetCacheFlushEnable
=
5329 flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
;
5330 pc
.DepthCacheFlushEnable
= flags
& PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
5331 pc
.StateCacheInvalidationEnable
=
5332 flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
5333 pc
.VFCacheInvalidationEnable
= flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
;
5334 pc
.ConstantCacheInvalidationEnable
=
5335 flags
& PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
5336 pc
.PostSyncOperation
= flags_to_post_sync_op(flags
);
5337 pc
.DepthStallEnable
= flags
& PIPE_CONTROL_DEPTH_STALL
;
5338 pc
.InstructionCacheInvalidateEnable
=
5339 flags
& PIPE_CONTROL_INSTRUCTION_INVALIDATE
;
5340 pc
.NotifyEnable
= flags
& PIPE_CONTROL_NOTIFY_ENABLE
;
5341 pc
.IndirectStatePointersDisable
=
5342 flags
& PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
;
5343 pc
.TextureCacheInvalidationEnable
=
5344 flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
5345 pc
.Address
= rw_bo(bo
, offset
);
5346 pc
.ImmediateData
= imm
;
5351 genX(init_state
)(struct iris_context
*ice
)
5353 struct pipe_context
*ctx
= &ice
->ctx
;
5354 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
5356 ctx
->create_blend_state
= iris_create_blend_state
;
5357 ctx
->create_depth_stencil_alpha_state
= iris_create_zsa_state
;
5358 ctx
->create_rasterizer_state
= iris_create_rasterizer_state
;
5359 ctx
->create_sampler_state
= iris_create_sampler_state
;
5360 ctx
->create_sampler_view
= iris_create_sampler_view
;
5361 ctx
->create_surface
= iris_create_surface
;
5362 ctx
->create_vertex_elements_state
= iris_create_vertex_elements
;
5363 ctx
->bind_blend_state
= iris_bind_blend_state
;
5364 ctx
->bind_depth_stencil_alpha_state
= iris_bind_zsa_state
;
5365 ctx
->bind_sampler_states
= iris_bind_sampler_states
;
5366 ctx
->bind_rasterizer_state
= iris_bind_rasterizer_state
;
5367 ctx
->bind_vertex_elements_state
= iris_bind_vertex_elements_state
;
5368 ctx
->delete_blend_state
= iris_delete_state
;
5369 ctx
->delete_depth_stencil_alpha_state
= iris_delete_state
;
5370 ctx
->delete_rasterizer_state
= iris_delete_state
;
5371 ctx
->delete_sampler_state
= iris_delete_state
;
5372 ctx
->delete_vertex_elements_state
= iris_delete_state
;
5373 ctx
->set_blend_color
= iris_set_blend_color
;
5374 ctx
->set_clip_state
= iris_set_clip_state
;
5375 ctx
->set_constant_buffer
= iris_set_constant_buffer
;
5376 ctx
->set_shader_buffers
= iris_set_shader_buffers
;
5377 ctx
->set_shader_images
= iris_set_shader_images
;
5378 ctx
->set_sampler_views
= iris_set_sampler_views
;
5379 ctx
->set_tess_state
= iris_set_tess_state
;
5380 ctx
->set_framebuffer_state
= iris_set_framebuffer_state
;
5381 ctx
->set_polygon_stipple
= iris_set_polygon_stipple
;
5382 ctx
->set_sample_mask
= iris_set_sample_mask
;
5383 ctx
->set_scissor_states
= iris_set_scissor_states
;
5384 ctx
->set_stencil_ref
= iris_set_stencil_ref
;
5385 ctx
->set_vertex_buffers
= iris_set_vertex_buffers
;
5386 ctx
->set_viewport_states
= iris_set_viewport_states
;
5387 ctx
->sampler_view_destroy
= iris_sampler_view_destroy
;
5388 ctx
->surface_destroy
= iris_surface_destroy
;
5389 ctx
->draw_vbo
= iris_draw_vbo
;
5390 ctx
->launch_grid
= iris_launch_grid
;
5391 ctx
->create_stream_output_target
= iris_create_stream_output_target
;
5392 ctx
->stream_output_target_destroy
= iris_stream_output_target_destroy
;
5393 ctx
->set_stream_output_targets
= iris_set_stream_output_targets
;
5395 ice
->vtbl
.destroy_state
= iris_destroy_state
;
5396 ice
->vtbl
.init_render_context
= iris_init_render_context
;
5397 ice
->vtbl
.init_compute_context
= iris_init_compute_context
;
5398 ice
->vtbl
.upload_render_state
= iris_upload_render_state
;
5399 ice
->vtbl
.update_surface_base_address
= iris_update_surface_base_address
;
5400 ice
->vtbl
.upload_compute_state
= iris_upload_compute_state
;
5401 ice
->vtbl
.emit_raw_pipe_control
= iris_emit_raw_pipe_control
;
5402 ice
->vtbl
.load_register_reg32
= iris_load_register_reg32
;
5403 ice
->vtbl
.load_register_reg64
= iris_load_register_reg64
;
5404 ice
->vtbl
.load_register_imm32
= iris_load_register_imm32
;
5405 ice
->vtbl
.load_register_imm64
= iris_load_register_imm64
;
5406 ice
->vtbl
.load_register_mem32
= iris_load_register_mem32
;
5407 ice
->vtbl
.load_register_mem64
= iris_load_register_mem64
;
5408 ice
->vtbl
.store_register_mem32
= iris_store_register_mem32
;
5409 ice
->vtbl
.store_register_mem64
= iris_store_register_mem64
;
5410 ice
->vtbl
.store_data_imm32
= iris_store_data_imm32
;
5411 ice
->vtbl
.store_data_imm64
= iris_store_data_imm64
;
5412 ice
->vtbl
.copy_mem_mem
= iris_copy_mem_mem
;
5413 ice
->vtbl
.derived_program_state_size
= iris_derived_program_state_size
;
5414 ice
->vtbl
.store_derived_program_state
= iris_store_derived_program_state
;
5415 ice
->vtbl
.create_so_decl_list
= iris_create_so_decl_list
;
5416 ice
->vtbl
.populate_vs_key
= iris_populate_vs_key
;
5417 ice
->vtbl
.populate_tcs_key
= iris_populate_tcs_key
;
5418 ice
->vtbl
.populate_tes_key
= iris_populate_tes_key
;
5419 ice
->vtbl
.populate_gs_key
= iris_populate_gs_key
;
5420 ice
->vtbl
.populate_fs_key
= iris_populate_fs_key
;
5421 ice
->vtbl
.populate_cs_key
= iris_populate_cs_key
;
5423 ice
->state
.dirty
= ~0ull;
5425 ice
->state
.statistics_counters_enabled
= true;
5427 ice
->state
.sample_mask
= 0xffff;
5428 ice
->state
.num_viewports
= 1;
5429 ice
->state
.genx
= calloc(1, sizeof(struct iris_genx_state
));
5431 /* Make a 1x1x1 null surface for unbound textures */
5432 void *null_surf_map
=
5433 upload_state(ice
->state
.surface_uploader
, &ice
->state
.unbound_tex
,
5434 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
5435 isl_null_fill_state(&screen
->isl_dev
, null_surf_map
, isl_extent3d(1, 1, 1));
5436 ice
->state
.unbound_tex
.offset
+=
5437 iris_bo_offset_from_base_address(iris_resource_bo(ice
->state
.unbound_tex
.res
));
5439 /* Default all scissor rectangles to be empty regions. */
5440 for (int i
= 0; i
< IRIS_MAX_VIEWPORTS
; i
++) {
5441 ice
->state
.scissors
[i
] = (struct pipe_scissor_state
) {
5442 .minx
= 1, .maxx
= 0, .miny
= 1, .maxy
= 0,